The fastest LLM on Earth.
Frontier level intelligence, 15x faster.
Benchmarks
A fraction of the latency.
Same tasks, same scoring. Celeris lands within a few points of the frontier on accuracy - and at more than 10x the speed.
01 · Intelligence
Figure 1. MMLU-Pro accuracy (5-shot, chain-of-thought) vs. response time (log scale). Up and to the left is better. The reasoning budget for Celeris, GPT-5, GPT-5 mini, and the Gemini models was set to 0, and Mercury 2 ran in instant mode - the fastest possible configuration for every model.
MMLU-Pro accuracy75.9
Response time · p50158 ms
Speedup vs GPT-513×
Output tok/s · p501,664
Speedup vs GPT-524×
Server time vs GPT 525×
02 · Speed
Figure 2. Output speed (p50 tokens per second) on ~1,000-token long-form prompts - 50 requests per model, outputs capped at 1,024 tokens. Full methodology →
Architecture
A new architecture for language model inference.
Current autoregressive models generate one token at a time. Every token depends on the one before it, making latency fundamentally sequential. Celeris is a new inference architecture for language models, built on diffusion, that achieves latency and quality previous diffusion systems could not.
The result is an OpenAI-compatible API that delivers intelligent responses in milliseconds.
For builders
Three lines to swap. Then everything is faster.
- ›The API is OpenAI-compatible. Point your SDK at inference.celeris.ai and keep the code you already wrote.
- ›Streaming is on by default. Responses as low as 24 ms, with no buffering or batch delay.
- ›Pricing is per token. You are billed for the tokens you generate, so speed costs you nothing extra.
quickstart.pypython
from openai import OpenAI client = OpenAI( base_url="https://inference.celeris.ai/celeris-1/v1", api_key="ck_…", ) # same call shape you already wrote stream = client.chat.completions.create( model="celeris-1", messages=[{"role": "user", "content": "Explain a Kalman filter"}], stream=True, )
Start building with Celeris.
Experience the new paradigm of inference.