Models

Melious runs open-weight models only. We don't host Claude, GPT-4o, or Gemini proper. The API shapes of OpenAI and Anthropic work anyway, because they're the shapes everyone writes clients against — but the inference itself runs on weights that are public.

The stance

Open-weight is a constraint, not a preference. It means:

• Every model we serve has publicly downloadable weights. Qwen, GLM, DeepSeek, Kimi, Mistral, Llama, FLUX, Whisper, Voxtral, and the rest.
• If a model is closed-weight, we don't serve it. There's no secret back door to a proprietary API.
• We map Anthropic/OpenAI model names where it helps unmodified clients keep working. Claude Code sends claude-sonnet-*; we map that to an open-weight model before inference. The response carries the original name back so the client doesn't get confused. See From Anthropic.

Why this shape? Because verifiable inference matters when your data is the input. You can inspect the weights we run. You can download them and benchmark them yourself. You can audit what your prompts go through. That's not a line we'd be able to hold if closed models were in the mix.

Families and capabilities

We group models by family rather than by vendor — the capability differences matter more than the org that released them.

General-purpose chat. GLM (Z.ai), Qwen (Alibaba), DeepSeek, Mistral Small, Llama Instruct, Hermes (NousResearch), GPT-OSS (OpenAI's open-weight release). These are the everyday workhorses — pick by context window, price, and speed.

Reasoning / thinking. DeepSeek-R1, Kimi Thinking, Qwen Thinking variants. Longer responses, higher cost, better on multi-step problems. Slower per token because they produce explicit intermediate reasoning. Use reasoning_effort: "high" when you want them to think harder, "low" to keep it quick.

Vision. Qwen2.5-VL and Qwen3-VL, Pixtral, Gemma 3 and 4. They accept image_url content blocks alongside text. See Vision.

Coding-specialized. Qwen3-Coder, DevStral, some Kimi variants. Tuned on code corpora — pick one when your workload is mostly software.

Audio input. Voxtral. A chat model that takes audio as input alongside text; different from the /v1/audio/transcriptions endpoint, which returns a plain transcript.

Embeddings. BGE-M3, Qwen3-Embedding, Multilingual E5. Multilingual by default. See Embeddings.

Rerank. Reranker variants of the above families. Takes a query plus a document list, returns them reordered by relevance. See Rerank.

Images. FLUX (the dev variants, including the FLUX-2 Klein distillations) and Qwen Image. Text-to-image, no editing. See Images.

Speech-to-text. Whisper large-v3 and turbo variants. 50+ languages. See Audio.

Finding the right model

We deliberately don't keep a static model list in these docs. Any table we'd paste here would be out-of-date by the next release. The hub is the source of truth.

Context windows and limits

Context windows vary by model — some are 8k, some are 200k+. The _meta.context_length field on GET /v1/models/{id}?include_meta=true gives the exact number for the model you're about to call. Exceeding it yields INFERENCE_3207 — trim the prompt or pick a model with more headroom.

The same _meta block also includes _meta.pricing with the cheapest user-facing rate across providers, _meta.max_output_tokens for the model's output cap, and _meta.reasoning_type to tell you if the model supports reasoning.

Request body size is capped at roughly 10 MB. Batch file uploads go up to 105 MB; see Files.

What about caching

Two things get conflated under "prompt caching," and they're worth separating:

• Explicit cache markers — the Anthropic cache_control block or an OpenAI-style cache breakpoint you set in the request — aren't supported. We drop cache_control blocks silently. There's no field to mark a prefix as cacheable.
• Transparent prefix caching is real and user-visible. Several of our providers auto-cache identical prompt prefixes and report the hit back. You see it as usage.cached_tokens (and nested in usage.prompt_tokens_details.cached_tokens) on Chat completions, and we pass the cheaper cache-read price through to you — cached tokens bill at a lower rate than fresh input tokens. You don't opt in; it happens when the provider reuses a prefix.

So: no handle to mark a cache, but caching does happen, and the response tells you how much of your prompt was served from it. If you want explicit cache control, tell us.

Related

Response fields that describe which model actually ran (for clients sending mapped names): Messages. Picking an equivalent to an OpenAI model name: From OpenAI. Capability detection programmatically: Models reference.