Skip to content

Model provider matrix

The C07 real local Ollama probe on 2026-09-19 UTC listed eight models and returned OK. from gemma3:4b-ctx32k through the new HTTP client, including input and output token usage. No paid provider endpoint was contacted.

Core maintainers review official API and price changes at least weekly with the vendor-discovery intake. A newly observed version or provider remains unverified until its exact API and host contract is exercised.

Observed 2026-09-19 UTC. This matrix describes the synapse worker HTTP path. Local means a real localhost HTTP exchange against a protocol fixture; it does not mean the vendor service was contacted. Unverified means no authorised live call established vendor compatibility. Unsupported means this worker does not perform the feature. The separate OpenCode participant delegates model choice to OpenCode and does not establish any provider API claim here.

Provider Wire family Auth Model list Stream parse Tool calls Usage HTTP errors and rate headers Cancellation Live API
Ollama OpenAI-compatible local Local Local Local Local parse; execution unsupported Local Local Local request abort; server stop unverified Local model path verified in C03
OpenAI Chat Completions Local Local Local Local parse; execution unsupported Local Local Local request abort; server stop unverified Unverified
DeepSeek Chat Completions Local Unverified Local family parser Local family parser; execution unsupported Local family parser Local Local request abort; server stop unverified Unverified
OpenRouter Chat Completions gateway Local Unverified Local family parser Local family parser; execution unsupported Local family parser Local Local request abort; server stop unverified Unverified
Qwen OpenAI-compatible regional endpoint Local Unverified Unverified model-specific delta semantics Local family parser; execution unsupported Local family parser Local Local request abort; server stop unverified Unverified
Mistral Chat Completions Local Unverified Local family parser Local family parser; execution unsupported Local family parser Local Local request abort; server stop unverified Unverified
Anthropic Messages Local Local Local Local parse; execution unsupported Local Local Local request abort; server stop unverified Unverified
Google GenerateContent Local Local Local Local parse; execution unsupported Local Local Local request abort; server stop unverified Unverified
xAI Chat Completions Local Unverified Local family parser Local family parser; execution unsupported Local family parser Local Local request abort; server stop unverified Unverified
Moonshot Chat Completions Local Unverified Local family parser Local family parser; execution unsupported Local family parser Local Local request abort; server stop unverified Unverified

ProviderHTTPClient returns tool calls as data. It accepts only a list of tool-call objects in a non-stream provider response and reports malformed collections as invalid_shape without exposing provider content. It never runs tools, retries a turn, or executes model output. The synchronous worker asks for one non-stream completion; its lower-level client can request SSE and normalizes the buffered result. Cancellation checks before and during reads, bounded by the socket timeout; provider-side cancellation acknowledgement is unknown. Rate headers are returned in ProviderReply; the worker does not yet adapt its scheduling to them. A quoted API price never implies a current account quota.

The relevant documented surfaces are OpenAI Chat Completions, DeepSeek chat completion, OpenRouter chat, Qwen compatible API, Mistral chat completion, Anthropic Messages, Gemini API, xAI chat completion, and Moonshot API. Vendor endpoint and model availability can change; review the exact vendor documentation and conduct an authorised live test before marking a vendor/version supported.