# Models

Model catalog (source: `contracts/models.yaml`). Use `model_id` as the `model` field in API requests.

External-provider models are available only to organisations explicitly granted the `external_providers` flag; without it they are absent from your model listing.

| model_id | tier | context (max_model_len) | publisher | released | license | description |
|---|---|---|---|---|---|---|
| `clef-flash` | utility | 16384 | Cloudflare | 2026-10-01 | apache-2.0 | Typed decisions (noul / choice / score) with calibrated probabilities; SystemOne request shape; text only. |
| `deepseek-ocr-2` | vision | 8192 | DeepSeek | 2026-01-27 | apache-2.0 | DeepSeek-OCR 2 ~3B — OCR / doc→Markdown (VLM, Visual Causal Flow, Apache-2.0). |
| `flux-2-klein` | image | 2048 | Black Forest Labs | 2026-01-15 | apache-2.0 | FLUX.2-klein 4B — high-end image generation (Black Forest Labs), detailed rendering faithful to the prompt, open license. |
| `gemma-4-26b-a4b` | large | 262144 | Google DeepMind | 2026-04-02 | apache-2.0 | Gemma 4 26B-A4B Instruct (QAT-AWQ) — MoE 26B/4B active, quality. |
| `gpt-oss-120b` | reasoning | 131072 | OpenAI | 2025-08-05 | apache-2.0 | gpt-oss 120B — served by OVHcloud AI Endpoints (external, EU). |
| `gpt-oss-20b` | reasoning | 131072 | OpenAI | 2025-08-05 | apache-2.0 | gpt-oss 20B (MXFP4) — MoE reasoning model. |
| `hypernova-60b` | reasoning | 131072 | Multiverse Computing | 2026-04-30 | apache-2.0 | HyperNova 60B (MXFP4 MoE) — Multiverse Computing's compression of gpt-oss-120b, 128k context. *Compressed from OpenAI gpt-oss-120b (Apache-2.0) by Multiverse Computing.* |
| `kokoro-82m` | audio | 4096 | hexgrad | 2025-01-27 | apache-2.0 | Kokoro-82M — near-instant speech synthesis (TTS), natural EN voices (US/GB), WAV 24 kHz, Apache-2.0. FR coming soon (espeak). |
| `kyutai-tts-1.6b-en-fr` | audio | 4096 | Kyutai | 2025-07-03 | cc-by-4.0 | Kyutai TTS 1.6B — English and French speech synthesis over complete WAV and incremental PCM WebSocket. *Kyutai TTS 1.6B weights CC BY 4.0; allowlisted Kyutai voice recordings CC0.* |
| `laguna-xs-2.1` | reasoning | 131072 | poolside | 2026-07-02 | openmdw-1.1 | Laguna-XS 2.1 (INT4 MoE) — poolside agentic coding model, 128k context. |
| `mistral-small-24b` | large | 32768 | Mistral AI | 2025-01-30 | apache-2.0 | Mistral Small 24B Instruct (AWQ) — best quality. |
| `nemotron3-nano-30b` | reasoning | 262144 | NVIDIA | 2025-12-04 | nvidia-open-model-license | NVIDIA Nemotron 3 Nano 30B-A3B (INT4 hybrid Mamba-2/MoE) — reasoning, 256k context. *NVIDIA Nemotron. Served under the NVIDIA Open Model License.* |
| `nemotron3-omni-30b` | vision | 131072 | NVIDIA | 2026-04-27 | nvidia-open-model-license | NVIDIA Nemotron 3 Nano Omni 30B-A3B (INT4) — image, video and audio chat, 128k context. *NVIDIA Nemotron. Served under the NVIDIA Open Model License.* |
| `qwen3-embedding-8b` | embedding | 7168 | Alibaba Qwen | 2025-06-05 | apache-2.0 | Qwen3 Embedding 8B — top-tier multilingual embeddings (among the best open models), adjustable dimension 256-4096, for demanding search. |
| `qwen3-vl-30b-a3b` | vision | 32768 | Alibaba Qwen | 2025-10-03 | apache-2.0 | Qwen3-VL 30B-A3B Instruct (AWQ MoE) — vision + text chat VLM. |
| `qwen3.6-35b-a3b` | reasoning | 262144 | Alibaba Qwen | 2026-04-16 | apache-2.0 | Qwen3.6 35B-A3B (AWQ) — MoE reasoning/coding, 35B/3B active. Long-context 256k. |
| `qwen3.8-27b` | reasoning | 65536 | Alibaba Qwen | 2026-08-14 | apache-2.0 | Qwen3.8 27B (INT4 W4A16) — dense reasoning/coding VLM, image + video, 64k window. |
| `voxtral-mini-3b` | audio | 8192 | Mistral AI | 2025-07-15 | apache-2.0 | Voxtral Mini 3B — audio transcription + understanding (Q&A, summarization), Apache-2.0. |
| `whisper-large-v3-turbo` | audio | 448 | OpenAI | 2024-10-01 | mit | Whisper large-v3-turbo — multilingual transcription (ASR), ~99 languages, MIT. |