Documentation · Reference
Models and limits
Tenvor serves one model today. Its limits and prices change, so this page tells you where to read them, not what they are.
The model
- API id
chat-35b-moe- Model
- Qwen 3.6 35B, with open weights (Apache-2.0)
- Next
- Qwen 4 27B, at launch
Send the id exactly as written. GET /v1/models lists the ids your key can use. The Models page has the full description.
GET /v1/models returns object:"list" and data containing only served models. GET /v1/models/{id} returns one card. Cards contain id, object, created, owned_by, context_length, max_model_len, max_output_tokens, pricing, architecture, supported_parameters and tenvor. pricing gives decimal USD per token as strings. tenvor contains schema_version, capabilities, limits, pricing_micro_usd_per_mtok and tenvor.reasoning. The reasoning object gives supported, default, parameter, output_field, levels, aliases, budget_share_of_max_tokens, min_max_tokens, billing and usage_reporting. Each level has id, label and budget_tokens; Max has a null budget and xhigh has api_only:true.
Where the live figures are
Models
The context window and the model's description.
Pricing
The price per million tokens in and out, and the minimum per request.
A program can read the same figures. This address needs no key:
curl https://orchestrator.tenvor.io/api/catalog/publicYou can also select the command and copy it manually.
For each model it gives max_context_tokens, max_input_bytes, max_output_tokens and the tariffs. A model that can think also lists its thinking levels and their limits in reasoning.
What the limits mean
The context window bounds input plus output. max_tokens above the safe output remainder is clamped. X-Tenvor-Output-Limit reports the requested and applied values. Input that fills the window is refused with context_length_exceeded. The input byte bound is max_input_bytes; a larger body returns 413 request_too_large.
Reused input
A client can ask a node to keep one conversation warm, so that a follow-up can reuse its working state. OpenCode's plugin and the Pi extension do; a plain OpenAI client does not, so it gets no reuse and no discount. Input the node reuses is billed at half the input price, and shows as cached_input_tokens on the receipt. Reuse is never guaranteed: the state lasts two minutes idle and fifteen minutes in total, and another request on that node erases it.
Content checked . Live limits and prices can change.