Changelog

Gemini Omni Flash 1.2 is Google's multimodal video model for generating, editing and extending clips with native audio, with text-to-video and image-to-video clips up to 15 seconds.

• Use google:gemini@omni-flash-1.2 to generate 3 to 15 second videos from text or a single starting frame, at 360p, 720p, 1080p or 4K in 16:9 and 9:16.
• Interpolates between a first and a last frame, and generates from up to 10 reference images and 3 reference videos in the same request, in clips of 3 to 10 seconds.
• Edits a source video of 3 to 15 seconds from a prompt when duration is left out, and takes reference images alongside it to bring in new elements.
• Extends a source video of up to 27 seconds by 3 to 10 seconds when duration is set.

GPT-Image-2.6 Flare and GPT-Image-2.6 Sunburst are now available for image generation and editing, with SVG and animated GIF output.

• Flare (openai:gpt-image@2.6-flare) is built for fast iteration in interactive workflows.
• Sunburst (openai:gpt-image@2.6-sunburst) is built for detailed compositions and precise editing.
• Both support text-to-image generation, prompt-guided editing, masks, and up to 16 reference images, with custom dimensions up to 3,840 pixels per side.
• inputs.markupImage takes a transparent PNG with marks drawn over the first reference image, like a circle around the object to change, as an alternative to inputs.maskImage.
• outputFormat adds SVG and GIF for text-to-image. A GIF has 2 to 16 frames (settings.gif.frameCount) of at least 0.02 seconds each (settings.gif.frameDuration), up to 960 pixels per side.
• settings.svg.model picks the model that writes the SVG, GPT-5.4 Mini by default, or GPT-5.4, GPT-5.5, GPT-5.6 Terra, GPT-6 Luna or GPT-6 Astra, and settings.svg.thinkingLevel sets its reasoning level.
• settings.thinkingLevel sets how much the model plans before rendering, low, medium or high (the default), independently of settings.quality, which goes from low to max.
• Transparent backgrounds are available with PNG or WebP output.

P-Video-2-Pro-Replace is Pruna AI's character replacement model for putting the person from one or two photos into an existing video, keeping its camera, motion and soundtrack.

• Use prunaai:p-video@2-pro-replace with a source video in inputs.video and one or two photos in inputs.referenceImages.
• Accepts source videos from 1 to 120 seconds and up to 8K, with aspect ratios from 1:4 to 4:1, and returns the same length and shape at 24 fps.
• Generates at 480p or 768p, or upscales to 1080p, 2K or 4K.
• Matches two photos to the people on screen from left to right, and takes an optional instruction of up to 4,000 characters to change that mapping or replace one person in a group.
• Keeps the source soundtrack and syncs the new person's lips to it, with settings.sourceAudioSync and settings.preserveAudio to turn either part off.

Hosted OpenClaw is deprecated and will be shut down on November 18, 2026. On that date, every hosted OpenClaw instance is deleted along with its workspace.

Kimi K3.1 Pro is now available through Runware's OpenAI-compatible layer.

• Use moonshotai:kimi@k3.1-pro as the model value in OpenAI-compatible requests.
• Supports text and image inputs, structured outputs, and tool calling for coding, 3D modeling, financial research, and agent workflows.
• Provides a 1,048,576-token context window with up to 1,048,576 output tokens.
• Reasoning is always on, with effort adjustable from low to max.

Kling VIDEO 4.0 Flash is now available on Runware as the 720p model of the Kling 4.0 family, covering text-to-video, image-to-video, reference-driven generation and prompt-based video editing.

• Use klingai:kling-video@4.0-flash to generate 3 to 20 second videos at 720p in 16:9, 1:1, 9:16 and 21:9.
• Starts from a first frame image, or from text alone.
• Accepts up to 10 reference images, 5 reference videos and 7 elements, and edits a source video from a prompt.
• Prompts point at each input by position, as <<<image_1>>> and <<<video_1>>>.

Krea 3 is Krea's image model for text-to-image generation and for editing selected regions of an existing image.

• Use krea:krea@3 to generate from text at 30 sizes from 1K to 4K, across ten aspect ratios from 1:1 to 21:9.
• Accepts up to eight reference images, with the output set by resolution (1K, 2K or 4K) or by one of the same 30 sizes.
• settings.editRegions takes up to 24 boxes. Boxes on the first reference image mark the areas to edit, and boxes on the other reference images mark the objects to bring in.
• Each box can carry a label, referenced from the prompt as @label, and a note with its own instruction.
• settings.thinkingLevel sets how much the model reasons before generating, fast or high.

The model router can now be held to a set of models you approve, with routing.allowedModels and routing.excludedModels.

• routing.allowedModels takes at least two distinct names and nothing outside the list is considered, so a request none of them can serve is refused.
• routing.excludedModels takes at least one name and every other model stays a candidate.
• Each name is an AIR identifier or a model slug, and the two lists cannot be sent together.
• A name the catalog does not publish is refused, with the name quoted in the message, so a typo cannot quietly shrink the set it was meant to describe.
• Names are matched case-insensitively, so bfl:2@1 and BFL:2@1 are the same model and a list holding both is refused as a repeat.

See the model router docs.

Topaz Labs Wonder 4 is a generative image enhancement and upscaling model that restores and enlarges an image in a single pass.

• Use topazlabs:wonder@4 with the upscale task type to upscale an image by a factor of 1 to 6, where 1 enhances it at its original size.
• Combines denoising, deblurring, compression artifact removal and face recovery in one pass, with no settings to tune.
• Produces outputs up to 256 megapixels and 32,000 pixels per side, so the largest accepted input shrinks as the factor grows.

HeyGen Video 1.0 generates clips up to 30 seconds long and can set the length from the dialogue in the prompt.

• Set duration up to 30 seconds for text-to-video and for image-to-video from a first frame, up from 15.
• Requests with reference images, videos or audio, or with settings.promptEnhancement set to disabled or ultraquality, keep the 15-second maximum.
• Set duration to auto to size the clip from the quoted dialogue in the prompt, at least 8 seconds.

MiniMax H3 Highspeed and MiniMax H3 Lite are now available as faster and lower-cost variants in the H3 video model family.

• H3 Highspeed (minimax:h3@highspeed) balances generation quality and speed for everyday content creation.
• H3 Lite (minimax:h3@lite) is the lightweight variant for concept previews, creative drafts and batch variations.
• Both generate 4–15 second videos at 480p or 768p across landscape, portrait, square and ultrawide aspect ratios, from text or from a first frame with an optional last frame.
• Both accept up to nine reference images, three reference videos and three reference audio clips for reference-driven generation.

Ideogram 4.5 Precise Edit can limit an edit to one region of the source with settings.editRegion, so sources larger than 4 megapixels keep their full resolution.

• Set settings.editRegion to a box of [x1, y1, x2, y2] pixel coordinates in inputs.seedImage. The box must lie inside the image, with each side at least 256 pixels, an aspect ratio between 1:6 and 6:1 and at most 4,194,304 pixels.
• Set it to auto to fit the region around inputs.maskImage, which is then required. A mask sent with a box must sit inside it.
• Everything outside the region comes back unchanged and the output keeps the source dimensions. Without a region, sources over 4,194,304 pixels come back scaled down.
• inputs.seedImage and inputs.maskImage now accept files up to 50 MB, and inputs.seedImage takes up to 67,108,864 pixels, such as 8192×8192.

Ideogram Vector 1.0 is Ideogram's vector model for generating SVG files from a prompt and converting existing images into SVG.

• Use ideogram:vector@1.0 with the vectorize task type, sending either positivePrompt to generate an SVG or inputs.image to convert one.
• Generates SVGs in 15 sizes from 1:3 to 3:1, up to 2048 pixels on the long side, at low, medium or high quality.
• Accepts up to 5 reference images to guide generation, and follows the first image's aspect ratio when no size is set.
• Converts an image at its own dimensions, scaled down to 2048 pixels on the long side, with a simple, detailed or automatic level of detail.

Jev 1.14 is now available on /v1/decisions, which takes TypeSafe's decisions format.

• Use typesafe:jev@1.14 as the model value in decisions requests.
• Answers typed choice, score, and yes-or-no questions with confidence, with stronger multilingual handling.
• Provides a 262,144-token context window for long logs, documents, and conversation histories in a single decision.

P-Video-Upscale is Pruna AI's video upscaling model for raising an existing clip to a higher resolution while keeping its aspect ratio and audio track.

• Use prunaai:p-video@upscale with the upscale task type to upscale a source video to 720p, 1080p, 2K or 4K.
• Keeps the source frame rate by default, or sets an output rate from 16 to 60 fps, interpolating frames above the source rate.
• Accepts source videos up to 512 MiB, with length capped by processed frames, from 2,400 at 720p to 300 at 4K.
• Includes a turbo mode for faster processing and an optional safety filter.

A task in a sync request over HTTP that is still running after 120 seconds now continues as an async task and comes back with status: "processing", instead of failing with a timeout error.

• Task polling with the same taskUUID returns its result or its error.
• Results that finished within the 120 seconds still arrive in the sync response, and getResponse returns the complete set.
• The TypeScript and Python SDKs poll for these tasks automatically from version 1.8.1.
• Video tasks accept "deliveryMethod": "sync".

P-Video-2-Pro-Avatar is Pruna AI's talking-video model for turning a single portrait into a lip-synced video driven by an audio track or a script.

• Use prunaai:p-video@2-pro-avatar with a portrait in inputs.image and either inputs.audio or speech.text.
• Matches the video length to the speech, from 1 to 180 seconds of audio, at 480p or 768p, or upscaled to 1080p, 2K or 4K.
• Reads scripts of up to 5,000 characters with 30 voices across 10 languages, and takes delivery instructions for tone and pacing.
• Accepts a prompt to steer motion and expression, with optional prompt upsampling.

GPT-6.1 Terra is now available through Runware's OpenAI-compatible layer.

• Use openai:gpt@6.1-terra as the model value in OpenAI-compatible requests.
• Supports text and image inputs, structured outputs, function calling, and tool use.
• Provides a 1,050,000-token context window with up to 128,000 output tokens.
• Configurable reasoning from none through max supports high-volume automation, visual reasoning, and computer-use agents.

gpt-oss-120b is deprecated and will be removed from Runware on October 30, 2026. After that date, requests using openai:gpt-oss@120b will return an error.

Kimi K2.6 is deprecated and will be removed from Runware on October 30, 2026. After that date, requests using moonshotai:kimi@k2.6 will return an error.

Kling VIDEO 4.0 is now available on Runware as a single model that covers text-to-video, image-to-video, reference-driven generation and prompt-based video editing.

• Use klingai:kling-video@4.0 to generate 3 to 30 second videos at 720p or 1080p in 16:9, 1:1, 9:16 and 21:9.
• Follows up to 10 keyframes per video, set as first, last and intermediate frames.
• Accepts up to 10 reference images, 5 reference videos and 7 elements, and edits a source video from a prompt.
• Prompts point at each input by position, as <<<image_1>>>, <<<video_1>>> and <<<element_1>>>.

HeyGen Voice 1.0 is now available on Runware for text-to-speech, in three variants that take the same parameters: heygen:voice@1.0, the faster heygen:voice@1.0-turbo, and heygen:voice@1.0-turbo-free, which runs Turbo on shared capacity.

• Speaks up to 5,000 characters of plain text from speech.text in 39 languages, detecting the language from the text when speech.language is not set.
• Uses a public voice from speech.voice, or clones a new one from a recording in inputs.audio, using its first 3 minutes.
• Returns a cloned voice as outputs.voiceId, which later requests from the same organization can send as speech.voice.
• Sets how expressive the delivery is with settings.expressiveness, from 0 to 1.

LTX-2.3 adds video extension, generating new seconds at the start or end of a clip sent in inputs.video.

• Set duration to the seconds to add, from 2 to 20. The output keeps the whole clip and comes back at 1920×1080, or 1080×1920 when the clip is taller than it is wide.
• settings.extend.mode adds the new seconds at the end, the default, or at the start.
• settings.extend.context sets how many seconds of the clip the model reads, from 1 to 20 and no longer than the clip. context plus duration must fit in 505 frames at the clip's frame rate, about 21 seconds at 24 fps.
• The clip must be an MP4 or MOV of at least 3 seconds and 73 frames, up to 3840×2160 and 100 MB.
• Extension requests take no width, height, fps, seed, steps, CFGScale, lora, negativePrompt, inputs.frameImages or inputs.audio.

Bria FIBO 2.0 and Bria FIBO 2.0 Edit replace FIBO and FIBO Edit. Requests keep using bria:20@1 and bria:21@1.

DeepSeek V4.1 Flash 1017 is now available through Runware's OpenAI-compatible layer.

• Use deepseek:v4.1@flash-1017 as the model value in OpenAI-compatible requests.
• October 2026 checkpoint of DeepSeek V4.1 Flash, which stays available as deepseek:v4.1@flash.
• Supports text and image inputs for coding, reasoning, visual understanding, and tool-using agent workflows.
• Provides a 1M-token context window with up to 1M output tokens and continuously adjustable reasoning effort.

MiniMax H3 Fast adds 768p output to its 480p tier.

• Set resolution to 768p when generating from a first frame or from first and last frames.
• For text-to-video and reference-driven generation, set width and height to 1344×768, 1024×768, 768×768, 768×1024, 768×1344 or 1792×768.
• 768p comes in the same six aspect ratios as 480p, from 21:9 to 9:16.

MiniMax M3.1 Flash is now available through Runware's OpenAI-compatible layer.

• Use minimax:m3.1@flash as the model value in OpenAI-compatible requests.
• Supports text, image, and video inputs for coding, tool use, agentic reasoning, and visual understanding.
• Provides a 1M-token context window with up to 524,288 output tokens.
• Reasoning is always on, with depth adjustable from low to max.

HeyGen Video 1.0 adds 1080p and 2K output to its 480p and 768p tiers.

• Set resolution to 1080p or 2k when generating from a first frame or from reference media.
• For text-to-video, set width and height to 1890×1080 or 1080×1890 for 1080p, and 2856×1632 or 1632×2856 for 2K.
• Both new tiers come in 16:9 and 9:16. 480p and 768p keep their six aspect ratios from 21:9 to 9:16.

FLUX 3 Image Klein is Black Forest Labs' open-weights image model for generation and reference-based editing, available on Runware as a distilled checkpoint and an undistilled base.

• Use bfl:flux@3-image-klein for the distilled checkpoint, built for speed, and bfl:flux@3-image-klein-base for the full undistilled checkpoint, the one to fine-tune and customize.
• Both generate from text and edit from up to 4 reference images in a single endpoint.
• Output width and height range from 128 to 4096 pixels, in multiples of 16.
• Bounding boxes place, move, resize or remove elements on a 0 to 1000 scale, each one referenced from the prompt by its id.
• Both take up to 50 steps, 16 by default on the distilled checkpoint and 50 on the base, plus a seed and four acceleration levels. The base checkpoint also takes a CFG scale and a negative prompt.

P-Video-2-Pro now generates from reference images, videos and audio, alongside text and first and last frames.

• Pass up to 9 images in inputs.referenceImages, counted together with any frame images, each up to 20 MB and 4096 pixels per side.
• Pass up to 3 clips in inputs.referenceVideos, each 2 to 15 seconds long and up to 100 MB, with 15 seconds in total.
• Pass up to 3 tracks in inputs.referenceAudios, each up to 50 MB, with 15 seconds in total. Audio references need at least one image, video or frame input.
• Requests with references use the speed or cost recipe in settings.mode.

Gemini 4 Argon is now available through Runware's OpenAI-compatible layer.

• Use google:gemini@4-argon as the model value in OpenAI-compatible requests.
• Supports text, image, video, audio, and PDF inputs for multimodal understanding and text-generation workflows.
• Provides a 1,048,576-token context window with up to 32,768 output tokens and configurable thinking effort.
• Built for long-horizon software engineering, enterprise knowledge work, and cybersecurity defense.

FFmpeg is now available on Runware as the first tool in the new tools task type, for processing video, audio and image files with raw FFmpeg options.

• Use runware:ffmpeg@0 to run one FFmpeg command over up to 10 input files passed by URL, each up to 5 GB.
• Accepts raw input options per file, plus output options, a filter graph and stream maps for the output.
• Returns a single file as MP4, WEBM, MOV or MKV video, MP3, WAV, AAC, FLAC or OGG audio, or a PNG, JPG or WEBP frame.
• Supports sync and async delivery.

Sending "model": "auto" hands the choice of model to the model router, which picks from what the request has to produce instead of from a name you pin.

• Available on imageInference, videoInference, audioInference, textInference, 3dInference, promptEnhance, caption, upscale, removeBackground, vectorize, imageMasking and controlNetPreprocess.
• Order the candidates with routing.preferences, which takes a quality, speed and cost weight from 0 to 5, where 0 leaves that dimension out of the decision.
• Cap the spend with routing.maxPrice, in USD, measured against what the request you sent would cost rather than against a headline rate. A model whose cost cannot be bounded is not chosen while it is set.
• State what the result has to meet in routing.requirements, up to 3,000 characters of plain words, and a model that cannot meet them is not chosen.
• Every parameter you send is kept as sent, and a model that cannot accept one of them is not chosen. When nothing can serve the request, the response names what ruled the rest out.

See the model router docs.

Gemini 3.1 Flash TTS is deprecated and will be removed from Runware on November 17, 2026.

No direct replacement is currently available in the Gemini TTS family on Runware. Update affected workflows before the deactivation date.

• Affected AIR ID: google:gemini@3.1-flash-tts

Gemini 3.7 Flash is deprecated and will be removed from Runware on November 12, 2026.

Google has replaced Gemini 3.7 Flash with Gemini 3.8 Flash, and requests using google:gemini@3.7-flash are already answered by Gemini 3.8 Flash. After November 12, 2026, requests using google:gemini@3.7-flash will return an error.

Qwen-Image-2.1-Turbo is the lower-cost tier of Alibaba's Qwen-Image-2.1 family for image generation and prompt-guided editing in a single endpoint.

• Use alibaba:qwen-image@2.1-turbo to generate from text, or edit and compose from up to 10 reference images.
• Prompt rewriting is on by default, with a direct or agent mode and optional thinking. The agent mode is available for text-to-image only.
• Output sizes range from 0.26 to 4.19 megapixels, at aspect ratios from 1:8 to 8:1.

GPT-5 Mini and GPT-5 Nano are deprecated and will be removed from Runware on December 11, 2026.

Nano Banana 2.1 is now available on Runware for image generation and editing with multi-reference grounding.

• Use google:nano-banana@2.1 for text-to-image generation and instruction-guided image editing.
• Supports up to 14 reference images and 10 reference videos for subject, product, and style consistency.
• Generates at 1K, 2K, and 4K, with minimal, medium, and high thinking levels.
• Supports web and image search grounding for generations based on current information and visual references.

Qwen-Image-2.1-Pro is the Pro tier of Alibaba's Qwen-Image-2.1 family for image generation and prompt-guided editing in a single endpoint.

• Generate from text, or edit and compose from up to 10 reference images.
• Prompt rewriting is on by default, with a direct or agent mode and optional thinking. The agent mode is available for text-to-image only.
• Output sizes range from 0.26 to 4.19 megapixels, at aspect ratios from 1:8 to 8:1.
• Supports negative prompts and seeds.

Grok Imagine Video 1.5 Lite is the lower-cost tier of xAI's Grok Imagine Video 1.5, for text-to-video and image-to-video generation with native audio.

• Generate video from a text prompt, or from a single image used as the first frame.
• Generate 1 to 15 second clips at 480p, 720p, or 1080p.
• Text-to-video takes width and height in seven aspect ratios from 16:9 to 9:16.
• Image-to-video takes resolution and follows the aspect ratio of the input image.

HeyGen Video 1.0 is HeyGen's video generation model for short clips with native audio.

• Generate video from a text prompt, from a first-frame image, or from reference media.
• Guide generation with up to 9 reference images, 3 reference videos, and 3 reference audios, with at most 12 references per request.
• Produces dialogue, ambience, and sound effects in the same generation, with no separate lip-sync step.
• Generate 5 to 15 second clips at 480p or 768p, in six aspect ratios from 21:9 to 9:16.

Ideogram 4.5 and Ideogram 4.5 Precise Edit are now available on Runware for image generation and instruction-guided editing.

Serverless is now available on Runware, letting you run your own model code on the same GPUs that serve the model catalog.

• Supports a Python codebase or a container image, and the platform builds it, scales it to demand, and routes requests to it.
• Scales workers up as traffic arrives and back to zero when idle, metered by the second while a worker is allocated.
• Validates every request against a schema taken from your own handler signatures, so a malformed body is rejected before a worker is woken.
• Works through the same API key, SDKs and CLI as the models API, with a task id you generate so a lost response can be retried without paying for the work twice.

See the Serverless docs to get started.

Transparent-background output for GPT Image 2 is deprecated and will stop working on September 30, 2026.

GPT Image 2 remains available for image generation and editing. To continue generating transparent backgrounds, migrate to GPT-Image-2.5 Flare or GPT-Image-2.5 Sunburst before the retirement date.

GPT-6.1 Sol is now available through Runware's OpenAI-compatible layer.

• Use openai:gpt@6.1-sol as the model value in OpenAI-compatible requests.
• Supports text and image inputs, structured outputs, function calling, and tool use.
• Provides a 1,050,000-token context window with up to 128,000 output tokens.
• Configurable reasoning from low through max supports long-running agents, document analysis, and multi-step automation.

P-Video-2 Pro adds a cost recipe to settings.mode, its lowest-cost way to generate a clip.

• Set settings.mode to cost, speed (the default) or quality.
• cost applies to text-to-video and to first-frame and first-and-last-frame generation, at 480p or 768p.

Recraft V4.1 Flash is the fast, low-cost tier of the Recraft V4.1 family for high-volume, design-focused image generation.

• Keeps output quality close to Recraft V4.1, with strong results in photography, portraits, complex scenes, and logo concepts.
• Generates 1K raster images at 14 sizes from 1:2 to 2:1, including 16:9 and 9:16.
• Omit width and height to let the model choose the aspect ratio from the prompt.
• Supports color palette and background color controls through settings.colors and settings.backgroundColor.

Boogu Image 0.1 Edit and Edit Turbo are now available on Runware for instruction-guided image editing.

• Edit a reference image with text instructions for object changes, background replacement, and style transfer.
• Preserve the input image's framing and generate at its native dimensions, up to 2048 pixels per side.
• Choose Edit for higher-fidelity output or Edit Turbo for faster iteration.
• Supports Chinese and English text rendering in edited images.

Gemini Omni Flash is deprecated and will be fully deactivated on September 30, 2026.

Migrate to Gemini Omni Flash 1.1 before the deactivation date. After September 30, 2026, requests using google:gemini@omni-flash will return an error.

Jev and Laya are now available on /v1/decisions, which takes TypeSafe's decisions format.

Claude models are now available on /v1/messages, which takes Anthropic's Messages format.

• Point the Anthropic SDKs at https://api.runware.ai and pass your Runware API key as the auth token.
• The request body reaches Anthropic unchanged, so every Messages parameter works, and streaming returns Anthropic's own events.
• model takes the AIR ID (anthropic:claude@opus-5.5), the Runware model ID, or Anthropic's model name.
• Claude Opus 5.5, Claude Fable 5.1 and Claude Sonnet 5 are new on Runware, each with a 1M-token context window, on both /v1/messages and /v1/chat/completions.

FLUX 3 Image is Black Forest Labs' native 4K model for image generation and editing on the FLUX 3 multimodal backbone.

• Supports text-to-image generation, image-to-image workflows, and prompt-guided local editing.
• Combines multiple references with identity and fine-detail preservation.
• Supports bounding-box placement for controlled composition and targeted changes.
• Renders in-place multilingual text across photorealistic, illustrative, and design-focused styles.

GPT Image 1 is deprecated and will be fully deactivated on October 23, 2026.

Migrate to GPT Image 2 before the deactivation date. After October 23, 2026, requests using openai:1@1 will return an error.

Grok Imagine Image Quality is deprecated and will be fully deactivated on November 2, 2026.

Migrate to Grok Imagine Image 2.0 before the deactivation date. After November 2, 2026, requests using xai:grok-imagine@image-quality will return an error.

GPT-6 Astra is OpenAI's frontier model for complex reasoning, coding, computer use, research, and document creation.

• Supports text and image inputs, structured outputs, function calling, and broad tool use.
• Provides a 1,050,000-token context window with up to 128,000 output tokens.
• Configurable reasoning from low through max supports latency and capability tradeoffs.
• Built for long-running agent workflows that plan, execute, and verify work across code, browsers, and professional software.

P-Video-2 Pro is Pruna AI's quality-tier video generation model for text and image-driven clips with native audio.

• Generate video from text or a first frame, with optional last-frame guidance for the final composition.
• Produces synchronized audio with support for two-shot dialogue and lip sync.
• Choose a speed or quality recipe, with off, turbo, or max prompt expansion.
• Generate 5 to 15 second clips at 480p or 768p and 24 FPS.

DeepSeek V4.1 Flash is now available through Runware's OpenAI-compatible layer.

• Use deepseek:v4.1@flash as the model value in OpenAI-compatible requests.
• Supports text and image inputs for coding, reasoning, visual understanding, and tool-using agent workflows.
• Provides a 1M-token context window with continuously adjustable reasoning effort.
• Uses a mixture-of-experts architecture designed to reduce memory and compute requirements for long-context workloads.

Seedance 1.5 Pro is deprecated and will be fully deactivated on November 11, 2026.

Migrate to Seedance 2.5 before the deactivation date. After November 11, 2026, requests using bytedance:seedance@1.5-pro will return an error.

FLUX Video Edit [fast] is Black Forest Labs' prompt-driven model for editing an existing video without rebuilding the whole clip.

• Use natural-language instructions to add, remove, or replace objects and characters, change settings or styles, and revise on-screen text or dialogue.
• Keeps unedited motion, camera work, timing, framing, and audio intact.
• Accepts source clips up to 15 seconds and preserves their duration and aspect ratio.
• Outputs edited video at 24 FPS and up to 720p.

P-Video 2 is Pruna AI's quality-focused video generation model, combining text, image, and audio-conditioned workflows in one model.

• Generate video from text, images, or an audio input, with native dialogue, music, and sound effects.
• Use optional first- and last-frame guidance to direct the opening and closing composition.
• Get sharper foreground subjects, stronger identity consistency, and improved lip synchronization for generated speech.
• Generate clips up to 20 seconds at 720p or 1080p, with 24 or 48 FPS output and a draft mode for iteration.

MiniMax H3 Fast and MiniMax H3 Max Turbo are now available as lower-cost, faster-iteration variants in the H3 video model family.

• H3 Max Turbo (minimax:h3@max-turbo) supports 480p and 768p video with native audio and optional first- and last-frame control, at half the per-second price of H3 Max.
• H3 Fast (minimax:h3@fast) generates 480p video for rapid previews and batch workflows.
• Both support text-to-video and image-to-video generation for clips up to 15 seconds with native synchronized audio.
• H3 Fast also accepts up to nine image references and up to three video or audio references for multimodal reference-driven generation.

Claude Opus 5 is now available through Runware's OpenAI-compatible layer.

• Use anthropic:claude@opus-5 as the model value in OpenAI-compatible requests.
• Supports text and image inputs for complex coding, reasoning, and document workflows.
• Provides a 1M-token context window and configurable effort levels for balancing capability, latency, and cost.
• Built for long-horizon agents and multi-step tool-use workflows that require sustained execution.

Gemini 3.8 Flash is now available through Runware's OpenAI-compatible layer.

• Use google:gemini@3.8-flash as the model value in OpenAI-compatible requests.
• Supports text, image, video, and audio inputs for multimodal understanding and text-generation workflows.
• Provides a 1M-token context window with configurable thinking effort for complex tasks.
• Designed for long-horizon coding, tool use, and autonomous agent workflows at Flash-class efficiency.

GPT-Image-2.5 Flare and GPT-Image-2.5 Sunburst are now available for image generation and editing.

• Flare (openai:gpt-image@2.5-flare) is the faster tier for high-volume generation and iteration.
• Sunburst (openai:gpt-image@2.5-sunburst) is the premium tier for detailed, iterative editing and tighter visual control.
• Both support text-to-image generation, prompt-guided editing, masks, and up to 16 reference images.
• Both support transparent backgrounds with PNG or WebP output, configurable quality, and custom dimensions up to 3,840 pixels per side.

Google is deprecating Nano Banana, also known as Gemini 2.5 Flash Image. The model will be fully deactivated on October 2, 2026.

If you are using Nano Banana, migrate to Nano Banana 2 before the deactivation date. After October 2, 2026, requests using google:4@1 will return an error.

P-Video-Edit is Pruna AI's instruction-based video editing model for changing a source clip while retaining its underlying motion, audio, and scene structure.

• Edits a source video from a text instruction, supporting attribute, object, and environment changes.
• Accepts up to four reference images to guide identity, style, and other visual details.
• Preserves the source clip's aspect ratio and duration, supporting videos up to 15 seconds at 25 fps.
• Includes a draft mode for faster, lower-cost previews before final-quality rendering.

Topaz Labs Proteus is now available through Runware for general-purpose video enhancement and upscaling.

• Restores detail while reducing noise and compression artifacts across a range of source footage.
• Provides adjustable controls for sharpness, detail recovery, deblurring, noise reduction, halo reduction, and grain.
• Supports focus correction for blurred subjects and preserves original detail where needed.
• Includes controls for progressive and interlaced source video, including field order selection.

Muse Image is Meta's image generation and editing model for reference-driven composition, structured layouts, and prompt-faithful visual work.

• Supports text-to-image generation and prompt-guided image editing in one endpoint.
• Accepts up to 10 reference images for multi-reference composition and style guidance.
• Generates 2K images in eight supported aspect ratios, from square to ultrawide.
• Can use image search, web search, and code execution to ground layouts, charts, tables, and other structured visual outputs.

MiniMax H3 Max is a performance-tuned video generation model for lower-latency text-to-video and image-to-video workflows.

• Supports text-to-video and image-to-video generation with native synchronized audio.
• Uses optional first- and last-frame images to control motion between keyframes.
• Generates 5–15 second videos at 480p or 768p across landscape, portrait, square, and ultrawide aspect ratios.
• Provides stronger prompt adherence and audiovisual quality for rapid iteration and high-volume generation.

GLM-5.3 Flash is now available through Runware's OpenAI-compatible layer.

• Use zai:glm@5.3-flash as the model value in your OpenAI-compatible requests.
• Supports standard OpenAI-compatible chat completions integration through Runware's text inference endpoint.
• Intended for teams that want to use Z.ai's latest Flash model without changing their existing OpenAI client setup.
• Promotional pricing runs until September 9, 2026 at $0.075 per 1M input tokens and $0.25 per 1M output tokens.

Gemini Omni Flash 1.1 is Google's updated multimodal video model for higher-resolution generation, longer clip workflows, and stronger reference-driven control.

• Supports output at 360p, 720p, 1080p, and up to 4K resolution.
• Extends clips in 3-to-10-second increments, up to 30 seconds total runtime.
• Adds start-frame and end-frame control for more directed interpolation between key beats.
• Accepts reference videos to preserve character identity and overall visual style across generations.

Four Recraft V4 Styles models are now available for reference-driven style-consistent generation across raster and vector workflows.

Wan3.0 is Alibaba's all-in-one multimodal video model for longer, reference-heavy generation and editing workflows.

• Supports text-to-video, first- and last-frame image-to-video, reference-driven generation, localized video editing, and temporal extension in one model.
• Accepts image, video, audio, document, and webpage inputs, with multimodal references that can be combined in a single request.
• Supports durations from 2 to 30 seconds, or auto to fit the attached content and prompt.
• Generates native synchronized audio by default, with prompt rewriting enabled by default for higher-quality generations.

Reve 2.1 (reve:2@1) is deprecated and will be fully deactivated on August 28, 2026.

If you still depend on this AIR ID in existing integrations, please migrate before the deactivation date or contact Runware for migration guidance.

KlingAI 1.0, KlingAI 1.5, KlingAI 1.6, KlingAI 2.0 Master, KlingAI 2.1, and KlingAI 2.1 Master video variants are deprecated and will be fully deactivated on September 15, 2026.

If you are still using any of these models, please migrate to a supported newer KlingAI model before the deactivation date.

Topaz Labs Wonder 3.5 is a generative image upscaler for recovering detail from soft, compressed, or undersized source images.

• Rebuilds fine detail in skin, hair, fabric, and surface textures rather than only enlarging existing pixels.
• Improves precision on structured content such as tables, charts, and dense text-heavy graphics.
• Supports low, medium, and high enhancement strength control to tune how aggressively detail is reconstructed.

Grok Imagine Image 2.0 is xAI's image generation and editing model with a single endpoint for both creation and prompt-guided edits.

• Supports text-to-image and image editing in the same endpoint.
• Generates outputs up to 2K resolution.
• Supports reference image input for guided edits and subject control.
• Supports 13 aspect ratios from portrait through ultrawide.

LTX-2.5 Fast and LTX-2.5 Pro are multimodal video generation models with native audio support and first-to-last-frame guidance.

• Both models support text-to-video and image-to-video workflows, plus an audio-driven path that pairs inputs.audio with a single inputs.referenceImages subject image.
• inputs.frameImages can drive first-frame and optional last-frame guidance for more directed motion and shot planning.
• Fast is the speed-focused tier, with output from 720p through 4K, support for 24, 25, 48, and 50 fps, and longer clip support at 720p and 1080p.
• Pro is the higher-capability production tier, with support for 24, 25, and 50 fps, output up to 1080p, and clips up to 10 seconds.

Lightricks LTX-2 Fast and LTX-2 Pro are deprecated and will be fully deactivated on August 15, 2026.

If you are still using either model, please migrate to the recommended replacement before the deactivation date.

During the transition window, requests targeting LTX-2 continue to resolve to LTX-2.3 equivalents and remain billed at LTX-2 rates. After August 15, 2026, requests to the deprecated models will return an error.

Seedance 2.5 is ByteDance's flagship video model for longer-form multimodal generation and editing workflows.

• Supports single continuous clips up to 30 seconds in one generation.
• Supports text-to-video, image-to-video, reference-driven generation, editing, and extension workflows.
• Accepts up to 50 multimodal reference inputs, including images, audio, and video.
• Improves prompt adherence by 20% compared with Seedance 2.0.

FLUX 3 Video is Black Forest Labs' first video model, built for multimodal generation with synchronized audio.

• Supports text-to-video, image-to-video, and video continuation workflows.
• Generates strong physics, motion, and audio in the same output pass.
• Works across hyper-realistic, cinematic, animation, and motion design outputs.
• Supports clips up to 20 seconds at 1080p.

Qwen-Image-3.0 and Qwen-Image-3.0-Pro are unified generation and editing models built for layout-sensitive image work, stronger text rendering, and multilingual visual design.

• Both variants support text-to-image generation and prompt-guided image editing in a single endpoint.
• Qwen-Image-3.0 is the balanced standard tier for everyday creative production, with support for one to three reference images and reliable layout handling.
• Qwen-Image-3.0-Pro is the higher-control tier for denser layouts, image-within-image compositions, and stronger small-text rendering in posters, menus, storyboards, and interface mockups.
• The family is built for multilingual typography-heavy workflows while preserving high photorealistic detail in faces, skin, hair, and materials.

DeepSeek V4 Flash 0731 is now available on Runware for long-context agentic, coding, and reasoning workflows.

• Reaches an Intelligence Index of 50, up 10 points over the previous DeepSeek V4 Flash.
• Built to run longer multi-step agentic tasks from start to finish.
• Strong on coding, reasoning, and tool use across production assistant workflows.
• Supports a 1M token context window while targeting the intelligence-versus-cost Pareto frontier.

Grok Imagine Video 1.5 now supports broader reference-driven video generation, including audio-conditioned workflows and higher output presets.

• Adds referenceImages input support, with up to 7 reference images for more guided video generation.
• Adds referenceAudios input support, enabling audio-conditioned generations when paired with reference images.
• Expands output presets to include 1080p, alongside 480p and 720p.
• Accepts direct width and height inputs for valid 480p, 720p, and 1080p output dimensions.

MiniMax H3 is MiniMax's new open-weights flagship video model, built for reference-driven generation with native synchronized audio.

• Supports native synced audio generated in the same pass as the video.
• Uses multi-reference conditioning to keep subjects and scene identity more consistent across shots.
• Supports extending existing clips or audio segments into continued video generation.
• Designed for text-driven, reference-guided video workflows with stronger continuity across multi-shot outputs.

P-Image-Ideogram is a text-to-image model built for strong prompt following and reliable in-image typography.

• Supports text-to-image generation with strong typography and more legible embedded text.
• Offers four generation modes, from Very Low for faster runs to High for better output quality.
• Generates at 1K or 2K, with standard aspect ratios or custom dimensions up to 2560 pixels.
• Supports prompt upsampling and seed control for more repeatable generations.

Ideogram Object Remover is a mask-based image cleanup model for removing unwanted elements without requiring a text prompt.

• Ranked #1 by FID on the independent OmniEraser benchmark.
• Takes an image and mask as input, with no prompt required.
• Removes objects, shadows, and reflections in a single pass.
• Supports multi-region removal within the same request.

LoRA training is now available on Runware for image workflows, with training and inference handled in the same platform.

• Supports five base models: FLUX.2 [klein] 4B, FLUX.2 [klein] 9B, FLUX.1 [dev], Qwen-Image, and Z-Image Base.
• Lets you train a LoRA and then run inference with it directly in the Runware ecosystem.
• Outputs safetensors files that are automatically imported into your account or available for download.
• Keeps trained outputs private by default, with ownership remaining with the account that created them.

Kimi K3 is coming to Runware on day zero as Moonshot AI's flagship multimodal reasoning model for coding, research, and long-horizon agent workflows.

• Open-weight model family built at 2.8T parameters for frontier-scale reasoning and software tasks.
• Supports a 1M-token context window for large codebases, document-heavy workflows, and sustained multi-step execution.
• Combines built-in reasoning with native multimodal understanding across text, images, and video.
• Get notified and start building as soon as the model goes live on Runware.

Reve 2.1 is Reve's flagship image model for native 4K generation and editing with stronger layout control and multilingual text rendering.

• Supports native 4K generation and editing across 17 aspect ratios.
• Uses layout-aware structure for denser scenes and more controlled complex compositions.
• Improves multilingual text rendering for typography-heavy and international design workflows.
• Supports up to 8 reference images for more guided generation, swaps, and composite-style editing.

BytePlus Video Enhancement is now available on Runware for upscaling, restoration, and higher-quality video finishing workflows.

• Upscales video up to 4K at up to 120 FPS for higher-resolution delivery and smoother playback.
• Supports frame interpolation and detail enhancement for cleaner motion and sharper output.
• Handles denoise, deblur, and SDR-to-HDR enhancement in the same restoration workflow.
• Standard is positioned for general-purpose content, while Pro is better suited to real-people footage with stronger skin and fine-detail recovery.

You can now query usage, spend, model performance, and error stats over the API, instead of only through the web console.

• Track requests and spend over a date range, broken down by day, model, or API key.
• Monitor per-model inference-time percentiles (average, p90, p99) to keep an eye on latency.
• Track client (4xx) and server (5xx) error counts and error rate, broken down by model or API key.
• Pull rolling usage totals (today, last 7 days, last 30 days, lifetime) alongside team, API key, and balance details in a single call.

See the Account Management docs for details.

Seedream 5.0 Pro is ByteDance's flagship image model for controlled generation, editing, and multi-reference workflows.

• Supports up to 10 reference images per generation for swaps, composites, and more guided visual control.
• Lets you set exact output dimensions directly, or describe the intended shape in the prompt for automatic aspect ratio selection.
• Supports 1K and 2K output targets for different production needs.
• Exports in PNG or JPEG for flexible design, marketing, and e-commerce workflows.

VEED Fabric 1.0 is a talking-video model for turning a single image and speech input into synchronized speaking footage.

• Generates talking videos from an image plus audio, or from image plus text-driven speech.
• Includes lip sync, head gestures, facial expression, and body motion in the generated clip.
• Infers the voice performance from the provided image-and-audio setup without a separate voice-cloning step.
• Works across photos, illustrations, and more stylized source artwork.

Gemini Omni Flash is Google's multimodal video generation and editing model for conversational video creation with native audio.

• Generates and edits 10-second video clips from text, photos, and video inputs.
• Supports photo-to-video workflows from up to 5 reference images.
• Adds video-to-video and multi-turn editing for iterative, conversational control.
• Combines native audio generation with Gemini-style world understanding for more guided video workflows.

Nano Banana 2 Lite is the lighter, faster image generation and editing variant in Google's Nano Banana 2 family.

• Built for lower-latency, higher-throughput image generation and editing workflows.
• Supports both text-to-image and image-to-image tasks.
• Keeps the same broad workflow shape as Nano Banana 2 while reducing model footprint.
• Good fit for teams that want faster turnaround without switching to a different image model family.

Seed Audio 1.0 is ByteDance's speech generation model for expressive text-to-speech, voice conversion, and speech editing workflows.

• Supports high-quality, controllable speech generation for narration, dubbing, and character voice use cases.
• Covers both text-to-audio and audio-to-audio workflows.
• Built for zero-shot and cross-lingual speech generation with strong speaker adaptation.
• Supports expressive control for emotion, voice conversion, speech editing, and speed editing tasks.

Fish Audio S2.1 Pro is a flagship multilingual text-to-speech model for expressive, low-latency speech generation.

• Supports 80+ languages with automatic language detection.
• Uses natural-language cues for emotion and delivery control.
• Supports multi-speaker dialogue in a single generation.
• Supports realtime streaming with very fast time to first audio.

Seedance 2.0 Mini is the lightest and most affordable model in the Seedance 2.0 family for faster multimodal video generation workflows.

• Compact Seedance 2.0 tier for lower-cost, higher-throughput generation while keeping the family’s multimodal video style.
• Supports multimodal reference input for more controlled generation workflows.
• Generates up to 15 seconds of video at 720p.
• Supports native audio generation.

HappyHorse 1.1 is Alibaba's upgraded multimodal video model for stronger motion, character consistency, and audio-visual synchronization.

• Reworked motion generation for more stable continuity and more expressive scene dynamics.
• Improved skin texture quality and facial detail, including better close-up performance.
• Stronger audio delivery for emotional dialogue with tighter lip-sync and audio-visual timing.
• Improved instruction following for longer prompt-driven storytelling, including up to 8 consecutive scenes per prompt.

Bria RMBG v2.0 (runware:110@1) is deprecated and will be fully deactivated on June 30, 2026.

This model has already been removed from the active catalog. If you still depend on this AIR ID in existing integrations, contact Runware for migration guidance.

OpenAI GPT Image 1 Mini and GPT Image 1.5 are deprecated and will be fully deactivated on December 1, 2026.

If you are using either model, please migrate to the recommended replacement before the deactivation date.

DeepSeek-V4-Pro is now available on Runware for long-context coding, reasoning, and agent workflows.

• Open-source frontier LLM positioned for state-of-the-art performance on agentic coding, math, STEM, and software tasks.
• Mixture-of-experts architecture with 1.6T total parameters and 49B active parameters.
• 1M token context window with up to 384K output tokens.
• Higher-capability DeepSeek V4 variant for stronger overall performance than Flash on demanding workflows.

Kling VIDEO 3.0 Turbo is a speed-optimized video generation model for faster, lower-cost production in the Kling 3.0 family.

• Faster and cheaper than standard Kling VIDEO 3.0 for high-volume workflows.
• Strong prompt adherence and improved multi-shot consistency across scene changes.
• Stable motion and improved lip sync for audio-aware video generation.
• Supports 720p and 1080p output.

Ray3.2 is Luma's flagship video model for controllable generation, transformation, and edit workflows.

• Supports text-to-video and image-to-video generation, including start and end frame control.
• Generates 5-second and 10-second clips at up to 1080p across multiple aspect ratios.
• Supports native HDR and EXR export for professional color-grading workflows.
• Includes looping, video extension, and broader video transformation workflows.

Claude Fable 5 is Anthropic's top-capability generally available Claude model for long-horizon coding, agentic execution, and multimodal reasoning.

• Supports text and image inputs in a single model.
• Built for extended coding, research, and tool-driven professional workflows with stronger sustained performance than earlier Opus models.
• Supports a 1M-token context window for large-context reasoning and long-running tasks.

FLUX Outpainting now supports a mode parameter to trade output quality for speed.

• high is the default mode and is recommended when fine detail, prompt adherence, or consistency with complex source content matters most.
• fast is significantly quicker for natural scene extensions such as landscapes, backgrounds, textures, and product images, with lower fidelity in the extended region than high and typically lower cost.

Runway Gen-3 Alpha Turbo and Runway Aleph are deprecated and will be fully deactivated on July 30, 2026.

If you are still using these models, please migrate before the deactivation date.

Runware is deprecating all ElevenLabs models currently available on the platform. These models will be fully deactivated on June 18, 2026.

If you are using any of these models, please migrate before the deactivation date. If you need help choosing an alternative on Runware, contact us for migration guidance.

Krea 2 Turbo is the fastest image generation variant in the Krea 2 family, built to preserve the same core creative workflow with much faster feedback.

• Delivers image generation in around 2 seconds for rapid prompt experimentation and early concept work.
• Works with style references, moodboards, and LoRAs, so it fits directly into existing Krea 2 workflows.
• Supports text-to-image and image-to-image workflows for fast visual iteration from prompts or source material.
• Better suited to exploration and option generation than final-image polish.

Riverflow 2.5 is a production-focused image family for brand, packaging, and commercial design workflows.

• Introduces controllable scoring-rubric generation for more structured creative evaluation and refinement.
• Supports transparent backgrounds, custom fonts, and outputs up to 4K resolution.
• Includes both text-to-image and image-to-image workflows for generation and editing.
• Pro is positioned for higher-accuracy production work, while Fast is optimized for lower-latency iteration.

P-Video-Replace is a character replacement model for existing video using a reference image to swap who appears on camera.

• Upload a video and reference image to replace the on-camera subject while keeping the original clip structure intact.
• Preserves motion, timing, camera behavior, lighting, and scene composition from the source video.
• Supports 720p and 1080p output.
• Pairs with P-Video-Animate for character-driven video workflows.

Ideogram 4.0 is a design-focused text-to-image model built for high-fidelity image generation with strong typography, structured layout control, and transparent background support.

• Strong multilingual text rendering, including dense small text and handwritten-style output.
• Supports natural-language prompts and structured JSON prompting for more reproducible design workflows.
• Layout-sensitive generation with bounding-box and composition control for posters, packaging, and branded graphics.
• Transparent background generation and 2K output for production-ready design assets.

MiniMax M3 is MiniMax's flagship multimodal model for coding, agentic execution, and long-horizon reasoning workflows.

• Supports text, image, and video understanding in a single model.
• 1M token context window for large codebases, research tasks, and extended multi-step sessions.
• Built for strong tool use, sustained agentic execution, and long-running software or browsing workflows.

Runway Aleph 2.0 is a video editing model for controlled changes to existing footage while preserving the rest of the clip.

• Edit one frame to define the change, then carry that edit through the relevant parts of the video.
• Applies edits across multiple shots and scene changes in a single pass.
• Preserves scene action, framing, and overall clip structure while changing only the requested elements.
• Supports clips up to 30 seconds at 1080p.

Exactly Illustrative Training is a custom model training workflow for building reusable brand-aligned illustrative models from a company dataset.

• Train a custom illustrative model on brand assets, campaign references, product visuals, packaging systems, and other proprietary illustration material.
• Produces a reusable trained model for consistent text-to-image and image-to-image generation on the platform.
• Built for teams that need stylistic consistency across campaigns, products, and internal creative workflows.

Grok Imagine Video 1.5 Preview is xAI's higher-tier image-to-video model for longer clips and stronger output quality.

• Generates image-to-video outputs from a single starting frame.
• Supports durations up to 15 seconds at 480p or 720p.
• Native audio support for music, sound effects, and dialogue.

Claude Opus 4.8 is Anthropic's highest-capability Claude model for advanced coding, agent systems, and long-context knowledge work.

• Supports text and image inputs in a single model.
• Built for demanding coding, agent orchestration, and multimodal reasoning workflows.
• Supports adaptive and extended thinking with a 1M-token context window.
• Designed for professional workflows that need strong instruction following and high-resolution vision support.

Grok 4.3 is xAI's flagship multimodal LLM for agentic reasoning, strong tool use, and low-hallucination workflows.

• Supports text and image inputs with a 1 million token context window.
• Configurable reasoning effort with none, low, medium, and high modes.
• Supports function calling and structured outputs for production assistants and agent-style systems.
• Built for coding, debugging, long-context analysis, and image-grounded reasoning workflows.

Nano Banana 2 now supports video input for context-aware image generation and editing workflows.

• Generate still images from video input with scene-aware visual context.
• Create thumbnails and infographics directly from footage.
• Maintains character and style consistency across frames.
• Supports output editing with a reference image and text prompt.

FLUX Virtual Try-On is a virtual try-on image editing model for transferring garments onto a person image.

• Accepts a person image plus one or more garment references to generate photorealistic try-on outputs.
• Works with packshots, model shots, and multi-garment compositions.
• Preserves the subject's face and pose while maintaining strong garment fidelity for logos, prints, stitching, and hardware.
• Built for fashion, e-commerce, and branded content workflows.

Two Krea 2 image models are now available for controllable generation and editing with weighted reference guidance.

FLUX Erase is a dedicated image editing model for clean, context-aware object removal.

• Removes unwanted objects, text, or distractions using mask-driven editing.
• Reconstructs surrounding content to match nearby structure, lighting, texture, and scene context.
• Built for image cleanup, privacy edits, and scene simplification without complex prompting.

DeepSeek-V4-Flash is a fast frontier language model for coding, reasoning, and long-context agent workflows.

• Supports both thinking and non-thinking modes for lower-latency and reasoning-heavy tasks.
• 1 million token context window with up to 384,000 output tokens.
• Supports tool calls and JSON output for structured API workflows and agent orchestration.
• Built for coding assistants, long-document analysis, large-repository reasoning, and cost-sensitive production LLM workloads.

FLUX Outpainting is a dedicated image expansion model for extending images beyond their original frame.

• Expands an existing image beyond its original borders in a single call.
• Continues scene content, lighting, texture, and composition naturally without requiring a text prompt.
• Supports exact target dimensions for aspect-ratio changes, banners, and social reformats.

ImagineArt 2.0 now supports prompt-guided image editing through the API.

• Edit images with text instructions and up to 4 reference images.
• Uses reasoning-based generation to interpret input images before producing edited outputs.
• Supports style transfer, lighting-aware scene changes, and background swaps.
• Edited outputs support up to 1.5K resolution across 15 aspect ratios.

Gemini 3.5 Flash is Google's frontier multimodal reasoning model for agentic and coding workflows.

• Supports text, image, video, audio, and PDF inputs in a single model.
• 1 million token input limit with up to 65,000 output tokens.
• Built for long-horizon workflows, sub-agent orchestration, and complex coding loops.
• Supports tool use including search, code execution, and function calling.

HeyGen Avatar V is a high-fidelity avatar video model for talking digital twins and other eligible registered avatar looks.

• Improves identity preservation, lip sync accuracy, facial expressiveness, and motion coherence across angle changes and long-form videos.
• Supports 500 pre-built avatars for studio-quality avatar video generation.
• Suited for marketing, training, sales, education, and localization workflows where avatar stability matters.

Veo 3.1 Fast now supports reference images for more consistent subject, product, and character shots.

• Pass reference images to guide subject, product, or character consistency.
• Generates up to 8 seconds at up to 2160p (4K) in 16:9 or 9:16 formats.
• Optional native audio support.
• Built for product shots, ads, and branded content workflows.

Four Recraft V4.1 models are now available for design-focused image generation, editing, and commercial asset production.

GPT-5.5 is OpenAI's frontier reasoning model for complex coding, long-context work, and tool-using professional tasks.

• Supports text and image inputs in a single model.
• 1,050,000 token context window with up to 128,000 output tokens.
• Configurable reasoning effort with support for web search, file search, code interpreter, hosted shell, apply patch, MCP, and computer use.

Kimi K2.6 is Moonshot AI's flagship multimodal LLM for coding, reasoning, and agentic execution.

• Supports text, image, and video understanding in a single model.
• Built for long-horizon software tasks, reliable tool use, and autonomous multi-step workflows.
• Designed for coordinated agent workflows and multimodal reasoning across complex tasks.

Three Anthropic Claude models are now available on Runware.

Inworld Realtime TTS-2 is a low-latency text-to-speech model built for expressive realtime voice interaction.

• Supports 133 voices across 14 languages.
• Free-form voice direction allows natural control over delivery style.
• Audio-aware delivery carries tone and pacing forward within a session for more natural results.

Grok Imagine Image Quality is xAI's quality-focused image generation and editing model for higher realism and tighter prompt control.

• Strong text rendering with high realism in complex scenes.
• Handles charts, infographics, diagrams, and UI concepts with tighter prompt following.
• Supports editing existing images with text instructions.

xAI is deprecating Grok Imagine Image Pro. This model will be fully deactivated on May 15, 2026.

If you are using this model, please migrate to the recommended replacement before the deactivation date.

Runware now supports MFA for individual accounts and workspaces, alongside Social Login and Enterprise SSO options for access control.

• Social Login supports Google and GitHub, and can upgrade an existing password account when the email address matches.
• Enterprise SSO supports SAML and OIDC providers, guided setup for common providers, and workspace access restrictions based on the configured identity provider. Contact Sales if your organization needs access.
• MFA supports SMS and TOTP authenticator apps, and workspaces can require MFA for all users.
• MFA is not compatible with Enterprise SSO, since SSO providers can enforce MFA directly for a single login flow.

Two Luma UNI-1 image models are now available for image creation, precision editing, and multi-reference control.

Mirelo SFX 1.5 converts silent video into synchronized, context-aware sound effects.

• Adds footsteps, impacts, ambience, and cinematic effects based on the visual scene.
• Maintains tight audio sync even in fast or complex motion sequences.
• Built for fast, robust, high-quality video sound design workflows.

PixVerse Modify is a video-to-video editing model for prompt-driven changes to existing footage.

• Swap subjects, age characters, and add or remove objects.
• Change weather, lighting, or time of day, and replace on-screen text.
• Supports full-video style transfer while preserving the underlying structure of the source footage.

P-Video-Avatar is a portrait-driven avatar video model for script- or audio-controlled talking-head generation.

• Supports 30 voices across multiple languages.
• Generates avatar videos at 720p or 1080p with durations up to one minute.
• Avatar movement, background, and speaking tone can be controlled through prompting.

OpenAI is deprecating DALL·E 2 and DALL·E 3. These models will be fully deactivated on May 12, 2026.

If you are using these models, please migrate to the recommended replacement before the deactivation date.

Juggernaut Z is a fine-tuned image model built on Z Image Base for more presentation-ready output.

• Improves lighting, depth of field, skin texture, and demographic balance.
• Reduces prompting overhead for usable photorealistic results.
• Compatible with existing LoRAs and Z Image workflows.

HappyHorse-1.0 is a video generation model for short-form text-to-video and image-to-video workflows.

• Generates up to 15 seconds of video at up to 1080p resolution.
• Supports synchronized audio generation in a single pass.
• Available across 7 languages.

SkyReels V4 is a video generation and editing model built for precise, element-level control.

• Edit specific parts of a video while preserving overall motion and scene flow.
• Supports image-to-video workflows with character consistency across shots.
• Omni References can tag inputs to control exactly what gets changed or replaced in the output.
• Outputs up to 1080p at 32 FPS for videos up to 15 seconds.

Google Imagen 3, Imagen 3 Fast, Imagen 4 Fast, Imagen 4 Preview, and Imagen 4 Ultra are deprecated. These models will be fully deactivated on June 30, 2026. New and existing integrations should migrate to the recommended replacement model.

Two Kling 4K video models are now available for native 4K generation without upscaling or post-processing workarounds.

Seedance 2.0 now supports 1080p video generation for higher-resolution production workflows.

• Full HD output with sharper detail and cleaner motion.
• Suited for campaign, social, and trailer deliverables.
• Character consistency workflows can use reference images generated with Seedream 5.0 Lite.

Google is deprecating Veo 2, Veo 3, and Veo 3 Fast. These models will be fully deactivated on June 30, 2026.

If you are using these models, please migrate to the recommended replacements before the deactivation date.

GPT Image 2 is OpenAI's state-of-the-art image generation and editing model for fast, high-quality output.

• Supports text-to-image generation and image editing in a single model.
• Handles flexible image sizes and high-fidelity image inputs.
• Built for strong prompt adherence, readable embedded text, and layout-aware visual output.

Gemini 3.1 Flash TTS is a low-latency text-to-speech model for real-time conversation and narration.

• Supports single-speaker generation and two-speaker dialogue output.
• Prosody can be controlled with style tags such as [laughs] and [short pause], as well as temperature.
• Includes 30 voices with multilingual output.

Three Qwen3-TTS 1.7B models are now available for multilingual text-to-speech generation with different voice control workflows.

Tripo 3D v3.1 is a high-fidelity 3D generation model for production-ready assets from text prompts or images.

• Supports text-to-3D and image-to-3D workflows.
• Improved geometry precision and texture quality compared to earlier versions.
• Supports PBR materials and export in GLB, FBX, OBJ, and STL formats.

ImagineArt 2.0 is a reasoning-based text-to-image model for instruction-faithful generation across realistic and stylized outputs.

• Uses reasoning-based generation for more accurate and nuanced prompt following.
• Supports a wide range of styles, from photorealistic imagery to illustration and anime.
• New color pipeline improves vibrancy and true-to-life color rendering.

Two Hunyuan 3D 3.1 models are now available for high-fidelity and rapid 3D asset generation from text prompts or images.

Two MiniMax music models are now available for song generation and melody-preserving audio transformation.

PixVerse v6 is now available for end-to-end video generation with native audio output.

• Generate video with dialogue, sound effects, and music in a single pass.
• Up to 15 seconds at up to 1080p.
• Improved consistency for characters and smoother transitions across shots, with directed camera motion.
• Currently ranked #4 for Image-to-Video on Artificial Analysis.

sync-3 is a lip sync model for full-shot generation with native 4K and super resolution support.

• Processes the entire shot in one pass instead of chunking, improving consistency across frames.
• Handles close-ups, extreme face angles, partially obscured faces, and obstructed mouths.
• Preserves voice, tone, and emotional expression across 95+ languages.

Meshy-6 is a 3D generation model for cleaner geometry and production-oriented mesh control.

• Supports text-to-3D and image-to-3D generation.
• Configurable quad or triangle topology with polycount and symmetry control.
• Supports low-poly workflows, A/T pose control, and textured outputs with optional PBR materials.

Three new OpenAI GPT-5.4 multimodal models for advanced reasoning, coding workflows, and high-volume automation.

Rodin Gen-2 is a 3D generation model for production-ready assets from text prompts or multiple input images.

• Supports text-to-3D and image-to-3D workflows with up to 5 input images.
• Generates quad-dominant or raw meshes with configurable polygon counts.
• Supports PBR and shaded materials with export in GLB, FBX, OBJ, USDZ, and STL formats.

ACE-Step v1.5 is an open-source music generation model with voice cloning, lyric editing, and multilingual support.

• Generates full tracks from text prompts with support for voice cloning, remixing, and cover generation.
• Text-to-audio and audio-to-audio workflows across a wide range of genres.
• Compositions up to 10 minutes with support for over 50 languages.
• Turbo variant available for faster generation.

16 new Exactly.ai models for style-consistent image generation across illustration and photography.

• 10 illustrative styles and 6 photographic styles, each trained with a specific artist for consistent output across generations.
• Outputs stay visually coherent without drifting between styles across batches.
• Suited for creative agencies, marketing teams, and content platforms generating images at scale with a consistent visual identity.

Wan2.7 is Alibaba's most capable video model, supporting text-to-video, image-to-video, reference-to-video, and instruction-based video editing.

• Instruction-based video editing to add, remove, or swap elements, change backgrounds, and restyle scenes from text.
• Reuse motion sequences, camera movements, and visual styles from reference videos.
• Lock character appearance and voice across shots with up to 5 subject references and 40+ facial expressions.
• Output at 720p or 1080p, 30fps with native audio and auto-dubbing support.

P-Image Upscale is an AI-powered image upscaling model that enhances resolution while preserving detail and realism.

• Upscale images to a target resolution between 1 and 8 megapixels.
• Optional detail enhancement to sharpen textures and fine elements.
• Realism enhancement to improve photographic quality and depth.
• Sub-second inference for fast, scalable upscaling workflows.

OpenAI is deprecating Sora 2 and Sora 2 Pro. These models will be fully deactivated on September 24, 2026.

If you are using these models, please migrate to an alternative video generation model before the deactivation date.

Veo 3.1 Lite is the most cost-effective model in the Veo family, matching Veo 3.1 Fast speed at roughly half the price.

• Text-to-video, image-to-video, and reference-to-video with 4s, 6s, and 8s output durations.
• Native audio generation with dialogue, sound effects, and ambient audio synchronized with video.
• Supports 16:9 and 9:16 aspect ratios at up to 1080p resolution.

Wan2.7 Image is a unified image generation and editing model with improved human figures, text rendering, and color output.

• Sharper human figure generation with more realistic proportions and detail.
• Improved multilingual text rendering supporting up to 3,000 tokens in 12 languages.
• More vibrant, saturated color output compared to previous Wan versions.
• Supports text-to-image, image-to-image, and region-based editing in a single model.
• Pro variant available for higher-fidelity output.

Topaz Labs Starlight Precise 2.5 is a diffusion-based video upscaling model with high-fidelity detail preservation.

• Upscale video to 1080p or 4K with fine texture and edge retention other models soften.
• Full temporal consistency across frames with natural skin textures and faces.
• Suited for film restoration, content upscaling pipelines, and production-grade video enhancement.

FLUX.2 [klein] 9B KV is a KV-cache optimized variant of Klein 9B with up to 2.5x faster multi-reference editing.

• All capabilities of FLUX.2 [klein] 9B: sub-second generation, text-to-image, and multi-reference editing in a single model.
• KV-cache support delivers up to 2.5x faster inference for multi-reference editing workflows.
• Built for interactive apps and real-time editing pipelines.

xAI Text-to-Speech is an expressive, low-latency text-to-speech model with multilingual support.

• Natural speech synthesis with inline controls for pauses, laughter, whispers, and emphasis.
• 5 distinct voices (Ara, Eve, Leo, Rex, Sal) across 20+ languages with automatic language detection.
• Low-latency synthesis suited for voice agents, real-time applications, and multilingual content at scale.

Two HeyGen models for end-to-end AI video creation and photorealistic avatar generation.

Bytedance is deprecating Seedance 1.0 Lite and Seedream 3.0. These models will be fully deactivated on May 13, 2026, with resources being gradually reduced starting from March 23rd.

If you are using these models, please migrate to the recommended replacements before the deactivation date.

Two new 3D generation models for creating production-ready assets from single images.

Three Gemini LLM models from Google are now available on Runware.

MiniMax M2.7 is one of the first LLM models available on Runware, built for complex agentic workflows and real-world productivity.

Grok Imagine Image Pro is xAI's most capable image generation model with higher quality, greater prompt adherence, and finer detail.

• Photorealistic and stylized image generation from text prompts.
• Native 2K resolution output for print, product visuals, and professional creative pipelines.
• Multi-turn iterative refinement with natural language instructions.

Inworld TTS 1.5 is a low-latency text-to-speech model available in Mini and Max variants, built for real-time AI applications at scale.

• Production-grade latency with P90 time-to-first-audio under 250ms (Max) and under 130ms (Mini).
• 15+ languages supported.
• Ideal for conversational AI agents, live translation, interactive media, and accessibility tools.

Kling VIDEO 3.0 now supports Motion Control with facial identity preservation and emotion reproduction.

• Consistent facial identity preserved across angles and through occlusions.
• Complex emotion reproduction with high-fidelity restoration.

LTX-2.3 is a video generation model with a rebuilt VAE, stronger prompt handling, and improved audio-video sync.

• Rebuilt VAE for improved textures, edge detail, and on-screen text clarity.
• 4× larger text connector for complex multi-subject prompts with reduced drift.
• Frame rate options at 24, 25, 48, and 50fps.
• Significantly reduced audio artifacts with tighter audio-video sync.
• Fast variant available for lower-latency generation.

Qwen-Image-2.0 and Qwen-Image-2.0-Pro are unified text-to-image and image editing models with native 2K resolution.

• Unified generation and editing in a single model.
• Native 2K resolution with fine detail in skin, fabric, and architecture.
• Professional typography rendering supporting up to 1,000-token prompts for infographics, posters, slides, and comics.

Two new Kling image models for professional-grade generation and high-fidelity reference-based workflows.

MiniMax Speech 2.8 is a text-to-speech model with consistent tone across long-form scripts.

• Consistent voice quality across extended scripts without tonal drift.
• Low latency suitable for real-time applications and voice agents.
• Turbo variant available for faster, lower-cost generation.

Nano Banana 2 delivers improved reasoning, rendering, and layout control for text-to-image generation.

• Stronger prompt adherence with real-world scene reasoning and stable multi-subject composition.
• State-of-the-art text rendering for posters, UI, and packaging.
• Improved material and lighting realism.
• Up to 4K output.

P-Video is an image-to-video model optimized for fast iteration with a draft-to-final rendering workflow.

• Draft mode generates preview clips at lower cost, with seed reuse for full-quality final renders.
• Strong input-image consistency with reliable lip sync and native dialogue.
• High-quality subject and background rendering, including low-resolution asset animation.
• Up to 1080p resolution at 48fps.

Bria FIBO Edit Tools is a structured image editing model with optional mask-based targeting for localized changes.

• Preserves original content outside the edit region.
• Supports recoloring, relighting, restoration, image blending, seasonal effects, and sketch generation.

Seedream 5.0 Lite is a text-to-image and image-to-image model with native 4K output and integrated web search.

• High-detail image generation up to 4K resolution.
• Strong text rendering for posters, UI mockups, and branded assets.
• Consistent multi-image outputs with stable character and layout coherence across batches.
• Integrates live web search to reflect current events and trending topics in generations.

Kling VIDEO 3.0 Omni is now available on the platform in both Standard and Pro variants.

• Unified text-to-video, image-to-video, and video-to-video workflows.
• Generates HD video with synchronized native audio.
• Supports reference-driven generation for stronger subject control.
• Includes prompt-based video editing with strong temporal consistency to maintain subject stability across clips.

Recraft V4 and V4 Pro are professional-grade text-to-image models built for brand and campaign work.

• Photorealistic and illustration output with structured composition.
• Native 2K generation with style and color control.

Kling VIDEO 3.0 is available in Standard and Pro variants for cinematic multi-shot video generation with native audio.

• Text-to-video and image-to-video support.
• Multi-shot generation with manual or automatic scene planning.
• Cross-shot consistency for characters and voices across scene cuts.
• Multi-language audio with lip sync.
• Up to 15 seconds with professional-grade resolution and reduced artifacts.
• Native text rendering for signage, captions, and advertising.
• Audio available as an optional add-on for both variants.

LoRA support is now available for P-Image and P-Image-Edit.

• Apply trained or community LoRAs for custom styles, characters, and concepts.
• Compatible with any HuggingFace LoRA without additional setup.
• P-Image-Edit LoRA supports style transfers and targeted image transformations.

Wan2.6 Flash R2V generates video from short reference clips while preserving visual identity and voice.

• Automatic lip sync and voice replication from the reference clip.
• Available at 720p and 1080p with optional audio output.
• Flash variant optimized for faster inference.

Z-Image is an open-source 6B parameter diffusion transformer for text-to-image and image-to-image generation.

• High visual fidelity with strong prompt adherence across diverse styles.
• Foundation model for distilled variants including Z-Image Turbo.

Vidu Q3 Turbo is a faster variant of Vidu Q3 optimized for rapid iteration.

• Up to 16 seconds at native 1080p with first and last frame support.
• Native audio-video sync with dialogue, ambient sounds, and BGM generated in a single pass.
• Available at 540p, 720p, and 1080p output.

Two new Grok Imagine models for photorealistic image generation and short-form video.

Runway Gen-4.5 delivers unprecedented physical accuracy and expressive character generation.

• Realistic weight, momentum, and fluid dynamics.
• Expressive characters with nuanced emotions and natural gestures.
• Strong prompt adherence for complex camera choreography and scene composition.
• Text-to-video and image-to-video supported.

LTX-2 19B is a production-ready model with synchronized audio generation.

• Native 4K resolution at 50fps.
• Up to 20 seconds duration.
• Prompt enhancer enabled by default to improve video quality.

Reve Image is an all-in-one image model for generating, editing, and remixing visuals.

• Supports style variation and layout changes.
• Handles semantic edits without rebuilding from scratch.
• Designed for iterative visual workflows.

Riverflow 2.0 offers reference-based super resolution and custom font control.

• Reference-based super resolution with high accuracy for product and marketing images.
• Custom font control with reliable text rendering.
• Consistent outputs with images checked and corrected before return.

Remix capability allows you to refine and modify existing videos without regenerating from scratch.

• Preserves original structure and camera movement for precise control.
• Suitable for color grading, style transformation, and element replacement.

Vidu Q3 generates synchronized video and audio from text or images.

• Sound effects, dialogue, and subtitles in a single output — no stitching required.
• Multi-shot support with automatic shot planning or prompt-based control.
• Up to 16 seconds at 1080p with stable, vivid results.
• Works from text prompts or image inputs.

Draft mode allows you to validate prompt intent and motion logic before full video generation.

• Save up to 40% compared to normal 480p generation.
• Fix prompts before compute-heavy rendering.
• Iterate faster and cheaper.

PixVerse v5.6 delivers more reliable single-shot video generation with improved audio.

• More reliable single-shot image-to-video and text-to-video with audio.
• Improved multi-character lip sync and dialogue pacing.
• Cleaner and more stable motion.
• More natural audio and stronger text-to-video consistency.

FLUX.2 [klein] 9B is the final model in the FLUX.2 klein family.

• 4-step distilled image generation and editing.
• Built for sub-second inference.
• Text-to-image and editing handled in a single model.
• Optimized for interactive apps and latency-sensitive production use.

TwinFlow Z-Image Turbo is an image generation model optimized for fast inference at 4 steps.

• Supports text-to-image generation with high-quality results.
• Low latency for rapid iteration workflows.
• Available via API only.

Bria Fibo Edit is an image editing model built around structured control using image, mask, and JSON instructions.

• Edits applied exactly where specified.
• Supports targeted edits, generative fill, outpainting, and compositional changes.
• Controlled, reproducible edits with full commercial safety through Bria's licensed training data.

Wan 2.6 Flash is a faster, lower-cost variant of Wan 2.6 for image-to-video workflows.

• Designed for preview workflows and high-throughput creative pipelines.
• Fast turnaround without sacrificing quality.

Kandinsky 5 Image is an open-source model producing high-quality visuals from text.

• Supports text-to-image and image-to-image workflows with reference images.
• Optimized for clarity and strong text adherence.
• Diverse visual concepts.
• Available via API only.

PixVerse v5 Fast delivers the same quality as PixVerse v5, optimized for speed.

• Text-to-video and image-to-video support.
• Ideal for rapid iteration and previews where turnaround matters.

Wan2.6 Image offers strong prompt adherence with spatial structure preservation.

• Guide style, subject, or background using reference images.
• Supports text-to-image and image-to-image workflows.
• Multiple reference images supported.

Bria FIBO Lite is a lightweight, efficient text-to-image model designed for rapid image generation.

Veo 3.1 now supports new output formats for high-resolution production workflows.

• Native 9:16 output for Instagram and TikTok-ready clips.
• State-of-the-art upscaling to 1080p and 4K resolution.

Z-Image Turbo now supports Image-to-Image workflows.

• Upload an image, add your prompt, and get fast, high-quality transformations.

Three new FLUX.2 [klein] models for fast, flexible image generation and editing.

ImagineArt 1.5 Pro is an evolution of ImagineArt 1.5 with enhanced hyper-realism.

• Outputs native 4K images with sharper textures.
• More realistic lighting.
• Built for production-ready image generation.
• Ideal for posters, layouts, marketing visuals, and design-sensitive formats.

Seedance 1.5 Pro now supports 1080p video generation with synchronized audio.

• Generate 1080p video with synchronized audio support.
• Works directly from text or image inputs.
• Precise audio-visual timing with strong motion coherence.
• Expressive camera control.
• Advanced narrative prompt handling for short-form video creation.

Creatify Aurora v1 turns a single image and audio input into a realistic avatar video.

• Accurate lip sync with natural facial motion.
• Expressive performance generation.
• Suitable for advertising, music videos, virtual humans, dubbing, and localization.

Three new Qwen image models for editing, generation, and layer decomposition.

Kling VIDEO 2.6 with Motion Control handles complex and fast actions with reference-based character performance.

• Built to handle complex and fast actions.
• Drive a character's performance using a reference video.
• Generate up to 30 seconds of one-take motion with synced expressions and lip sync.