Services & modalities
inference.club is not just chat. A provider can expose many services, each declaring a type that decides which endpoints it answers and how requests are routed to it. A transcription request only ever lands on a speech-to-text service, even if two services happen to share a model name.
Which modalities are live at any moment depends on what the providers are running. The reference cluster (the project's own home GPU fleet) serves all of them; the model names below are what it runs — on the open network, call GET /v1/models to see what's actually reachable for you.
The catalog
| Service | Endpoint(s) | In → out | Reference model |
|---|---|---|---|
| Chat / LLM | /v1/chat/completions, /v1/completions | text → text | Qwen3, Nemotron |
| Speech-to-text | /v1/audio/transcriptions | audio → text | Nemotron ASR |
| Text-to-speech | /v1/audio/speech | text → audio | Magpie TTS |
| Voice cloning | /v1/voice/generations | script + samples → audio | Dia |
| Image | /v1/images/generations, /v1/images/edits | text/image → image | FLUX.2 Klein |
| Video | /v1/videos/generations | text/image → video | LTX-2 |
| Music | /v1/music/generations | text + lyrics → audio | ACE-Step |
| 3D mesh | /v1/3d/generations | image → 3D (GLB) | TRELLIS |
| Audio enhance | /v1/audio/enhance | audio → audio | Maxine Studio Voice |
| Scrape | /v1/scrape | URL → markdown | Firecrawl |
Text & language
The llm services speak the full OpenAI chat surface — streaming, multimodal messages (image inputs on vision models), tool calling, and JSON-mode extraction on models that declare those features. The legacy /v1/completions endpoint is there for older clients; new work should use chat completions.
Audio
Three distinct audio services. Speech-to-text transcribes uploaded audio and, when the model advertises the timestamps feature, returns word- and segment-level timings. Text-to-speech synthesizes speech from text with a selectable voice (GET /v1/audio/voices lists them). Voice cloning is a richer TTS path: you supply a [S1]/[S2] dialogue script and optionally map each speaker to a short audio sample from your library, and the engine generates a multi-speaker performance. Audio enhance takes a noisy recording and returns a cleaned one.
Vision & 3D
Image services generate and edit stills. Video generates clips from a text prompt or an initial image (image-to-video). 3D mesh turns a single image into a downloadable GLB model you can drop into a <model-viewer> or a game engine.
Music
Music generation takes a text description plus optional lyrics and structure hints (duration, BPM, key) and returns a finished track. It pairs naturally with image generation for cover art — the song-and-cover workflow template does exactly that.
Utility: scrape
Not every service is a GPU model. Scrape fetches a URL and returns clean Markdown — useful as the first step of a research workflow or as a tool the playground agent can call.
How a request finds a service
Every endpoint resolves a provider the same way: find the first online provider that serves the requested model and the matching service type, then proxy the call over the tailnet. See Routing for the selection rules and Architecture for the full path a request takes.
Most of these endpoints can run synchronously (you wait for the result) or asynchronously (you get a job id and poll). See Direct vs async.