What's new on Runware: October 2026

The Runware Sonic Pod Factory

On October 30, we announced a modular AI data center factory that will build 10 gigawatts of AI compute capacity a year. It turns the data center from a construction project into a manufactured product: complete Sonic Pods come off a single production line, like cars, at up to 30 megawatts of capacity a day.

Each Sonic Pod is a whole data center in a shipping container, with 1 megawatt in a 20-foot container or 2 megawatts in a 40-foot one. The cooling loops, power electronics, circuit boards, and network switches are designed and built in-house, and every pod is tested at full load before it ships. At a site that already has power, a pod goes from delivery to first token in 48 hours.

The line is configurable across NVIDIA's range, from RTX PRO 6000 for high-volume inference to GB300 and Vera Rubin, including NVIDIA's new DSX POD reference design. Production starts in Europe in the first quarter of 2027 and reaches the full 10-gigawatt rate by the end of that year.

The factory will also expand the infrastructure behind Runware's own inference platform. Read the announcement, see the factory page, or reserve capacity.

Runware Serverless

On October 2, we launched Runware Serverless. It runs your own models, code, or containers on the GPU infrastructure behind Runware's Model APIs, including our own Sonic Pods, while we take care of the queues, scaling, and monitoring. Workers run on 1, 2, 4, or 8 GPUs and can scale to zero when they're idle.

On-demand GPUs are billed by the second. The launch post covers how it works and what you can run on it.

Since the launch, Serverless has added:

• Multi-GPU checkpointing. GPU checkpointing now covers workloads that span several GPUs. A worker restores an already-initialized workload instead of repeating the expensive startup on every GPU, which cuts cold-start time for larger models and distributed inference.
• Environments. Keep development, staging, and production workloads apart, each with its own deployments and configuration, so you can test a change without touching production.
• OIDC connectors. Connect workloads to external services with OpenID Connect. A workload trades its trusted identity for short-lived credentials, so your deployments hold fewer long-lived secrets.
• Secrets templates. Set up secrets from a template and reuse the same configuration across deployments, with sensitive values kept out of your application code.
• Faster cold starts. The time from a request to a workload that's ready to run is down by about 4 seconds on average.
• One-click getting started. A shorter path from finding Serverless to deploying your first GPU workload.

Kling VIDEO 4.0

Kling VIDEO 4.0 is Kling's new flagship, and it puts everything in one model. Kling 3.0 came as separate Standard, Pro, Omni, and Turbo models; 4.0 covers text-to-video, image-to-video, reference-driven generation, and editing an existing video from a prompt.

It generates 1080p clips of up to 30 seconds and follows up to 10 keyframes per clip, including first, middle, and last frames. One request can combine reference images, reference videos, and reusable elements to keep subjects, style, and motion consistent. That makes it a strong fit for longer narrative shots, storyboard-driven sequences, and production work where a character or product has to look the same from shot to shot.

On Runware, Kling VIDEO 4.0 costs $0.157 per second at 720p and $0.235 per second at 1080p, so a 10-second 1080p clip is $2.35. With a video input, it's $0.188 per second at 720p and $0.282 at 1080p. See it in action on the Kling VIDEO 4.0 page.

FLUX 3 Image and FLUX 3 Image Klein

Built from 6 references

• 
  Face
• 
  Outfit
• 
  Snowcat
• 
  Motorcycle
• 
  Samoyed
• 
  Research station

Black Forest Labs opened the month with FLUX 3 Image, its first image model built on the multimodal FLUX 3 backbone it shares with FLUX 3 Video. It generates natively at up to 4K, composes from up to 10 reference images while keeping identities and fine details, places elements with bounding boxes, and edits text in place, with accurate typography in text-heavy layouts. Pricing is flat across text-to-image and editing, and only the resolution changes it: $0.041 per image at 0.75K, $0.048 at 1K, $0.10 at 2K, and $0.607 at 4K.

Two weeks later, BFL released the open weights. FLUX 3 Image Klein is the distilled checkpoint, built for speed, and FLUX 3 Image Klein Base is the full undistilled model to fine-tune and customize. Both handle text-to-image and reference-based editing. Klein is the one for interactive apps, rapid iteration, and high-volume jobs: a 1024×1024 image at 16 steps costs $0.006.

For four generations of FLUX side by side, read From FLUX.1 [dev] to FLUX 3 Image, or see more examples on the FLUX 3 Image page.

Nano Banana 2.1

Built from 2 references

• 
  Face
• 
  Smock

Google's update to Nano Banana 2 improves graphic composition and subject consistency, and follows complex prompts and edit instructions more closely. It works from text alone or from up to 14 reference images plus a reference video. It can also ground a generation in live web and image search, so the result reflects current facts and real visual references.

It renders at 1K, 2K, and 4K, with three thinking levels that trade speed for reasoning depth. That suits layout-heavy design work, recurring characters and products, and precise multi-step edits.

Nano Banana 2.1 is billed by tokens. In our test runs, a 1K image came to about $0.034 and a 2K image to about $0.052, roughly half of Nano Banana 2's $0.069 at 1K and $0.103 at 2K.

Also new this month

Image

Video

Audio

Text and LLMs

These models run through Runware's OpenAI-compatible API.

Decision models

Jev 1.14 runs on /v1/decisions, which takes TypeSafe's decisions format, rather than the OpenAI-compatible API. For what decision models are for, read Jev, Laya, and the emerging role of decision models.

Tools

Model updates

New features for models that were already on Runware.

• LTX-2.3: video extension. Add 2 to 20 seconds to the start or end of an existing clip.
• HeyGen Video 1.0: 1080p and 2K. Two new resolution tiers, in 16:9 and 9:16.
• MiniMax H3 Fast: 768p. A new tier alongside 480p, in the same six aspect ratios.
• P-Video-2-Pro: reference inputs. Generate from up to 9 reference images, 3 videos, and 3 audio tracks, alongside text and first and last frames.
• Ideogram 4.5 Precise Edit: region edits. Limit an edit to one region of the source, so images over 4 megapixels keep their full resolution.

Platform updates

• Long sync requests no longer time out. A task in a sync request over HTTP that's still running after 120 seconds now continues as an async task and returns status: processing instead of a timeout error. Poll with the same taskUUID for the result. The TypeScript and Python SDKs do this for you from version 1.8.1. Video tasks now accept "deliveryMethod": "sync" too.

Every change, with the parameters, is in the changelog.

Deprecations and migrations

Switch before a model's removal date. After it, requests to that model return an error. gpt-oss-120b and Kimi K2.6 were removed on October 30.

Model	Removal date	Switch to
gpt-oss-120b (openai:gpt-oss@120b)	Removed October 30, 2026	DeepSeek-V4.1-Flash (deepseek:v4.1@flash) or Gemma 4 31B (google:gemma@4-31b)
Kimi K2.6 (moonshotai:kimi@k2.6)	Removed October 30, 2026	Kimi K3 (moonshotai:kimi@k3)
Gemini 3.7 Flash (google:gemini@3.7-flash)	November 12, 2026	Gemini 3.8 Flash (google:gemini@3.8-flash), which already answers its requests
Gemini 3.1 Flash TTS (google:gemini@3.1-flash-tts)	November 17, 2026	No direct replacement yet
GPT-5 Mini (openai:gpt@5-mini)	December 11, 2026	GPT-5.4 Mini (openai:gpt@5.4-mini)
GPT-5 Nano (openai:gpt@5-nano)	December 11, 2026	GPT-5.4 Nano (openai:gpt@5.4-nano)

From the blog

• Runware Serverless: run your own AI from $0.63 per GPU-hour
• From FLUX.1 [dev] to FLUX 3 Image: four generations of Black Forest Labs models
• Runware unveils AI factory to build modular AI data centers just like cars