Model parameter matrix

The matrix uses canonical names. Protocol adapters can use a native spelling (max_output_tokens, stop_sequences, generationConfig.topP, and similar) when the live catalog maps it to the same parameter.

Parameter	Meaning
temperature	sampling temperature
top_p	nucleus sampling
max_tokens	output token limit
frequency_penalty	frequency penalty
presence_penalty	presence penalty
stop	stop sequences
seed	deterministic sampling hint
n	number of candidates
stream	streaming response
response_format	structured output format
tools	tool/function declarations
tool_choice	tool selection policy
logprobs	token log probabilities
top_logprobs	top log-probability candidates
logit_bias	token bias map
user	caller/user label

model, messages, input, contents, prompt, and their protocol-specific required equivalents are structural fields. They are not optional capability claims and are never rejected merely because a capability entry is absent.

Reading the live catalog

curl https://api.nexus-hub.ru/v1/catalog

Each verified entry has this shape:

{
  "endpoint_family": "chat.completions",
  "optional_parameters": ["temperature", "tools"],
  "verified_at": "2026-08-15T00:00:00Z"
}

An empty optional_parameters array means that the absence of all listed optional parameters was verified for that model and endpoint. An absent entry means that Nexus does not yet know and keeps compatible passthrough behaviour.

Safe request examples

The examples below are valid shapes only when the live catalog lists the corresponding capability for the selected model and endpoint.

Tools:

{"tools":[{"type":"function","function":{"name":"lookup","parameters":{"type":"object"}}}],"tool_choice":"auto"}

Structured output:

{"response_format":{"type":"json_object"}}

Penalties and streaming:

{"temperature":0.2,"top_p":0.9,"frequency_penalty":0.1,"presence_penalty":0,"stream":true}

Log probabilities:

{"logprobs":true,"top_logprobs":3}

Support can differ between OpenAI-compatible, Anthropic-compatible and Gemini-compatible routes, and can differ between models in the same family. The available models themselves remain key-specific: check GET /v1/models before selecting a model.