Celeris-1: a diffusion LLM that replies in 158ms, near GPT-5-level
Parallel diffusion, not token-by-token generation: 1,664 tok/s at $2/$6 per M on an OpenAI-compatible API — but 8K context caps it to short tasks.
Copy markdown
158ms replies, near-frontier scores
Celeris-1 answers with a p50 latency of 158ms and streams up to 1,664 output tokens/sec — roughly 24x faster than GPT-5 — while scoring 75.9% on MMLU-Pro, just under GPT-5-mini's 78%. Near-frontier intelligence now lands in the time a page paints.
Diffusion, not token-by-token
Instead of emitting one token at a time, Celeris-1 generates text in parallel with a diffusion architecture — the same bet Inception's Mercury made, but Celeris claims 5x the speed and about 12 points more on MMLU-Pro than Mercury 2. Diffusion LLMs are moving from demo to production-ready.
The catch: an 8K context window
Celeris-1 tops out at 8,192 tokens and Celeris tunes it for short, structured work, not long-form. Reach for it on classification, routing, autocomplete and single-turn tool calls — keep big-repo coding on a long-context model.
$2/$6 per M, drop-in via OpenAI API
It's live today at $2 per million input tokens and $6 per million output through a streaming, OpenAI-compatible endpoint — point your base URL and key at console.celeris.ai and go. Usage-based billing, no waitlist.
Build this weekend: kill the pause
Speed is the feature: wire it into a voice agent so replies land before the caller notices dead air, add inline autocomplete that keeps pace with typing, or run live classification on every keystroke. Projects that felt laggy at 20 tok/s feel instant past 1,600.