Models
Supported providers, model-id format contract, and capability notes.
Providers
The BYOK settings surface ships a 12-provider catalog
(frontend/src/lib/provider-models.ts, PROVIDERS): DeepSeek, OpenAI,
Anthropic, OpenRouter, Google AI, Groq, Together AI, Fireworks AI, DeepInfra,
xAI (Grok), Glenn Guilloux AI, plus OpenAI Compatible (Custom) — which
accepts an arbitrary base_url so any OpenAI-schema upstream works.
Model-id format contract (C4)
Model ids are type/model qualified, e.g. deepseek/deepseek-v4-flash in
the built-in catalog. Strict upstreams additionally require their own
namespace prefix in fully-qualified form — for example
deepseek-chat/deepseek-v4-flash. Sending a bare model name to such an
upstream can fail with HTTP 400.
Practical rule: pick models from the picker (which inserts correctly-formed ids), and when configuring custom providers copy the exact id spelling from that provider's docs.
Context windows
Selected defaults from the catalog (PROVIDER_MODELS, field maxTokens):
| Provider | Model | Max tokens |
|---|---|---|
| OpenAI | GPT-4o / GPT-4o Mini | 128 000 |
| Anthropic | Claude Sonnet 4 | 200 000 |
| Gemini 2.5 Pro | 1 048 576 | |
| DeepSeek | V4 Flash / V4 Pro | 8 192 |
| Groq | Llama 4 Scout | 131 072 |
The catalog evolves with the product; treat this table as representative, not
exhaustive. Reasoning-style models budget part of the window for internal
reasoning — size max_tokens requests accordingly.
Last updated 2026-08-25 (git-derived)