Compatibility matrix

A high-density at-a-glance view. Rows are the canonical endpoints the prober checks. Columns are the major OSS implementations.

Legend: ✅ pass · ⚠️ pass-with-deviation · ❌ not implemented · — out of scope

Core surface

Endpoint	OpenAI	vLLM	llama.cpp	Ollama	LM Studio	TabbyAPI
/v1/models	✅	✅	✅	✅	✅	✅
/v1/models/{id}	✅	✅	⚠️	✅	⚠️	✅
/v1/chat/completions	✅	✅	✅	✅	✅	✅
/v1/chat/completions (stream)	✅	✅	✅	✅	✅	✅
/v1/completions	✅	✅	✅	⚠️	✅	✅
/v1/embeddings	✅	✅	✅	✅	✅	❌
/v1/audio/transcriptions	✅	✅	❌	❌	❌	❌
/v1/audio/speech	✅	⚠️	❌	❌	❌	❌
/v1/audio/voices	✅	❌	❌	❌	❌	❌
/v1/images/generations	✅	⚠️	❌	❌	❌	❌
/v1/images/edits	✅	❌	❌	❌	❌	❌

Extensions

Endpoint	OpenAI	vLLM	llama.cpp	Ollama	LM Studio	TabbyAPI
/v1/responses	✅	⚠️	❌	❌	❌	❌
/v1/responses/compact	✅	❌	❌	❌	❌	❌
/v1/realtime (WebSocket)	✅	❌	❌	❌	❌	❌
/v1/audio/translations	✅	⚠️	❌	❌	❌	❌
/v1/moderations	✅	❌	❌	❌	❌	❌
/v1/files	✅	❌	❌	❌	❌	❌
/v1/batches	✅	❌	❌	❌	❌	❌
/v1/uploads	✅	❌	❌	❌	❌	❌
/v1/fine_tuning/jobs	✅	❌	❌	❌	❌	❌

Feature deviations on shared endpoints

For endpoints all servers implement, what's actually different:

Behavior	OpenAI	vLLM	llama.cpp	Ollama	LM Studio	TabbyAPI
Streaming usage when requested via stream_options.include_usage	✅	✅	✅	⚠️	✅	✅
Honors seed deterministically	✅	✅	✅	⚠️	⚠️	⚠️
response_format: json_schema	✅	✅	✅	⚠️	⚠️	⚠️
tool_calls[].function.arguments is a string	✅	✅	✅	⚠️	✅	✅
OpenAI error envelope on 4xx	✅	✅	⚠️	⚠️	⚠️	⚠️
cache_prompt / KV-cache reuse	n/a	✅	✅	⚠️	✅	✅
Logprobs on chat completions	✅	✅	✅	⚠️	⚠️	⚠️
Repetition penalty (repetition_penalty/repeat_penalty)	❌	✅	⚠️	⚠️	⚠️	✅

The repetition-penalty row is a field-name split, not just a support gap: vLLM, SGLang, and TabbyAPI take repetition_penalty; llama.cpp — and the llama.cpp-backed LM Studio and Ollama — take repeat_penalty. OpenAI has neither and uses frequency_penalty/presence_penalty. Unknown keys are dropped silently, so a mismatched field name is a no-op, not a 4xx. Details: Chat & completions → Sampling parameters.

Retired

Endpoints OpenAI used to expose but no longer route:

• /v1/images/variations — DALL·E-2-era image variation endpoint. An unauth probe of api.openai.com on 2026-05-16 returned 404; the public Images API docs no longer list it. The canonical image-variation flow is now /v1/images/edits with gpt-image-1. Pruned from the catalog in v0.3.1 (see issue #7).

Caveats

• Snapshots are best-effort. Each cell reflects the most recent aioc probe run we have a record of, against the latest stable release. Pre-release builds may differ.
• "Honors seed" is a Phase C implication test — currently manual. We mark it ⚠️ when reproducibility was inconsistent in our tests, even if the parameter is accepted.
• The matrix isn't a leaderboard. A red row may simply mean a server doesn't target that surface (Ollama isn't trying to do audio).

Help us keep it honest

PRs that update a cell with a fresh aioc probe report attached are the easiest way to contribute. See Contributing.