Capabilities

GET /api/models/capabilities is the live sheet behind the models cards. It is sourced from chain registration (productscience/inference/inference/models_all), not a spreadsheet.

Fields

Field	Meaning
`id`	Same id you send as `model`
`context_length`	Prompt + output window the host advertises
`max_tokens`	Typical max completion (often 16K)
`supports_tools`	Native or proxy-assisted tool calling
`tool_call_parser`	Host parser name, when set
`supports_reasoning`	Host can emit separated reasoning
`reasoning_parser`	Parser id
`quantization`	e.g. FP8
`v_ram_gb`	GPU memory the host registration claims
`hf_repo` / `hf_commit`	Hugging Face pin
`units_of_compute_per_token`	On-chain compute units
`validation_threshold`	Network validation floor

How to use it

Pick context for RAG / agents from context_length, not marketing names
If supports_tools is false, still try tools — the proxy can simulate — but prefer a tools-capable id
Quantization and VRAM are host facts, not quality SLAs

OpenAI list vs capabilities

GET /v1/models is the short list of currently routable ids. Capabilities may still mention a model that just unlisted. Always send an id that /v1/models returns right now.

Jev

jev is not a Gonka model. It is a real decision model, not a chat model. The calling guide, request shapes, and eligibility rules are on Jev. MiniMax, DeepSeek, GLM, Qwen, and Kimi on this API run on Gonka.