Models
Open models you can run on your own workers, each one deployable without writing any code.
Introduction
Each entry names a model and what serves it, and the build turns that into the source your app runs. You deploy it as it is, without writing any code.
You run the app on your own workers, and you pay for the hardware it holds rather than per request.
These are not the models the inference API serves. To call a model over an API without running anything yourself, use the inference API instead.
Gemma 4 12B
Google- Text generation
- l40s
- Apache-2.0
The mid-size Gemma, for when the 31B is more model than the work needs.
Gemma 4 31B
Google- Text generation
- rtx-pro-6000
- Apache-2.0
Google's instruction-tuned 31B, served on its own OpenAI-compatible route.
GLM-4.7 Flash
Z.ai- Text generation
- rtx-pro-6000
- MIT
A mixture-of-experts model that activates a fraction of its weights per token, so it answers at the speed of a much smaller one.
GPT-OSS 20B
OpenAI- Text generation
- l40s
- Apache-2.0
OpenAI's smaller open model, published already quantized, so 21B parameters fit the cheapest card.
Nemotron 3 Nano 30B A3B
NVIDIA- Text generation
- rtx-pro-6000
- NVIDIA Open Model License
A mixture-of-experts model with a hybrid attention and state-space architecture, activating a fraction of its weights per token.
Nemotron 3 Nano 4B
NVIDIA- Text generation
- l40s
- NVIDIA Open Model License
NVIDIA's small model, tuned for the hardware it runs on.
Nemotron 3.5 Lightning 30B A3B
NVIDIA- Text generation
- rtx-pro-6000
- NVIDIA Open Model License
The latency-tuned member of the Nemotron 3.5 line, a mixture of experts over a hybrid attention and state-space backbone.
Qwen3.5 4B
Alibaba- Text generation
- l40s
- Apache-2.0
Small enough to keep a worker warm without the bill that usually implies.
Qwen3.5 9B
Alibaba- Text generation
- l40s
- Apache-2.0
A mid-size model that answers quickly and leaves most of the card for the context.
Qwen3.6 35B A3B
Alibaba- Text generation
- rtx-pro-6000
- Apache-2.0
A mixture-of-experts model that activates a fraction of its weights per token, in FP8.
Qwen3.8 27B
Alibaba- Text generation
- l40s
- Apache-2.0
Alibaba's current flagship open model, in the FP8 weights its authors publish, which halve the memory and let it run on the cheapest card.
SmolLM3 3B
Hugging Face- Text generation
- l40s
- Apache-2.0
A 3B model trained for reasoning and multilingual use, small enough that a worker comes up in seconds.
BGE Base English
BAAI- Embeddings
- l40s
- MIT
Normalized 768-dimension English embeddings, over a 512-token context.
BGE Large English
BAAI- Embeddings
- l40s
- MIT
Normalized 1024-dimension English embeddings, over a 512-token context.
BGE Small English
BAAI- Embeddings
- l40s
- MIT
Normalized 384-dimension embeddings from a list of texts in one request, on the smallest card that fits it.
BGE-M3
BAAI- Embeddings
- l40s
- MIT
Multilingual embeddings across a hundred languages, for retrieval that is not only English.
Qwen3 Embedding 0.6B
Alibaba- Embeddings
- l40s
- Apache-2.0
Current-generation embeddings from a list of texts in one request, on a model small enough to cost almost nothing to keep warm.
Qwen3 Embedding 4B
Alibaba- Embeddings
- l40s
- Apache-2.0
2560-dimension embeddings over a 40k-token context, the middle size of the family.
Qwen3 Embedding 8B
Alibaba- Embeddings
- l40s
- Apache-2.0
4096-dimension embeddings over a 40k-token context, the largest of the family.
Clef 27B
Cloudflare- Decisions
- rtx-pro-6000
- Apache-2.0
Typed decisions over text, JSON and images. Send a state and your questions, and every option of every question comes back scored in one pass, with nothing to parse.
Clef Flash 9B
Cloudflare- Decisions
- l40s
- Apache-2.0
The small, quick Clef. The same typed decisions over text, JSON and images, on a third of the card.
Clef Omni 30B A3B
Cloudflare- Decisions
- rtx-pro-6000
- Apache-2.0
The Clef that also listens. Send a state with recordings and images alongside it, and your questions come back answered with a probability each.
Laya
Convai Innovations- Decisions
- l40s
- Apache-2.0
Typed decisions instead of written answers. Send a state and your questions, and each one comes back answered with a calibrated probability, in English or in a hundred other languages.
FLUX.2 dev
Black Forest Labs- Image generation
- rtx-pro-6000
- FLUX non-commercial license
The full FLUX.2, not distilled, so it takes a real schedule rather than four steps. Its 113 GB of bfloat16 weights are quantized to int8 as they load, which is what puts it on one card. Its weights are gated, so this one needs your own Hugging Face token stored as HF_TOKEN.
FLUX.2 Klein 4B
Black Forest Labs- Image generation
- l40s
- Apache-2.0
The small distilled FLUX, four steps to an image, Apache licensed and light enough to leave a worker warm.
FLUX.2 Klein 9B
Black Forest Labs- Image generation
- l40s
- FLUX.2 Klein license
The larger distilled FLUX, still four steps. Its weights are gated, so this one needs your own Hugging Face token stored as HF_TOKEN.
Ideogram 4
Ideogram- Image generation
- l40s
- Ideogram license
Ideogram's open release, in the NF4 weights its authors publish for diffusers. Known for typography, so text inside the image comes out readable. Its weights are gated, so this one needs your own Hugging Face token stored as HF_TOKEN.
Qwen Image 2.1
Alibaba- Image generation
- l40s
- Qwen license
Alibaba's open image model, strong on rendering text inside the image it generates.
Qwen Image 2.1 Turbo
Alibaba- Image generation
- l40s
- Qwen license
The distilled Qwen Image, a few steps instead of a full schedule, for when latency matters more than the last detail.
Z-Image Turbo
Alibaba- Image generation
- l40s
- Apache-2.0
A distilled image model that reaches a finished picture in single-digit steps, so a request costs a fraction of a full sampler run.
LTX-2.5
Lightricks- Video generation
- rtx-pro-6000
- LTX-2.5 license
Video with its own audio from one pass, and the model picks the length to suit the prompt. Its weights are gated, so this one needs your own Hugging Face token stored as HF_TOKEN.
MiniMax H3
MiniMax- Video generation
- rtx-pro-6000
- MiniMax license
Video with its own stereo audio from one denoising pass, at 768p. This is H3-Base, the part MiniMax published. The context stage their own workflow runs in front of it is not open, so results here are the base model's rather than what their API returns.
MiniMax Music 3
MiniMax- Audio generation
- l40s
- MiniMax Music 3 community license
A song from a description and its lyrics, sung, in stereo at 44.1 kHz. Leave the lyrics out and the track comes back instrumental.
YuE2 3B
M-A-P- Audio generation
- l40s
- CC BY-NC 4.0
A full song from a style and its lyrics, sung, in stereo at 48 kHz. It writes a score first and hands it back as ABC notation, so an edited score can be sent in to render a revision, or a transcribed melody to cover an existing song.
Qwen2.5-Omni 7B
Alibaba- Speech
- rtx-pro-6000
- Qwen license
A model that listens and answers, in writing or out loud. Send a question, or a recording with a question about it, and ask for the reply to be read aloud.
Qwen3-TTS 0.6B Base
Alibaba- Speech
- l40s
- Apache-2.0
Text read aloud in the voice of a recording you supply. A few seconds of clean speech is enough to copy from.
Qwen3-TTS 0.6B Custom Voice
Alibaba- Speech
- l40s
- Apache-2.0
Text read aloud in one of the voices the model publishes, at a size that starts in seconds.
Qwen3-TTS 1.7B Base
Alibaba- Speech
- l40s
- Apache-2.0
Text read aloud in the voice of a recording you supply, at the larger size for closer copies.
Qwen3-TTS 1.7B Custom Voice
Alibaba- Speech
- l40s
- Apache-2.0
Text read aloud in one of the voices the model publishes, with an optional instruction for how to read it.
Qwen3-TTS 1.7B Voice Design
Alibaba- Speech
- l40s
- Apache-2.0
A voice built from a description rather than picked from a list. Say the age, accent, pace and mood, and the model reads your text in it.
Voxtral Mini 3B
Mistral AI- Speech
- l40s
- Apache-2.0
Speech turned into text. Send a recording and get its transcript, with the language detected or named.
The deployed app
It behaves like any other Serverless app: it scales to zero on its idle TTL, resumes on the next call, and apps scale changes its hardware like anywhere else.
The weights are pinned to one commit. An entry records the revision it was tested against, so the same deploy brings the same weights however long it has been listed.
The server is the one the model's own ecosystem publishes. A text model runs under vLLM on its own OpenAI-compatible route, from the image vLLM's maintainers release, so there is no Runware layer between your request and the server.