celeris-1 generates via diffusion and claims a 157 ms median latency

Celeris Labs' celeris-1 model delivers a median latency of 158 ms. What this speed changes for classification and extraction tasks.

A language model that produces multiple tokens per invocation instead of just one per pass: this is the proposition from Celeris Labs, founded by Tom Hamer and Jesse Clark, previously the creators of Marqo. Its first model, celeris-1, replaces autoregressive decoding with an inference architecture based on diffusion.

The lab measures a median response time of 158 milliseconds and a throughput of 1,664 tokens per second, compared to 69 for GPT-5 under the same protocol, with an MMLU-Pro score of 75.9%. A second series of tests, conducted from EC2 instances co-located with each provider, places celeris-1 ahead of Cerebras by 8% and ahead of Groq by a factor of 3.5, while running on standard GPUs rented from AWS rather than dedicated silicon.

The phrase “frontier level intelligence” sparked reactions. The chart published by Celeris itself places GPT-5 at 81.9% and Gemini 3.5 Flash Lite at 83% on MMLU-Pro, therefore above the model, which several developers pointed out. Others are calling for a complete quality/latency frontier rather than an isolated p50, including high percentiles, long outputs, and tool-use tasks, arguing that failed generations disappear in a median. The size of the context window, deemed short, and the pricing, two dollars per million input tokens and six per output, fuel the same reservations.

Furthermore, the documentation steers celeris-1 toward short intermediate tasks such as classification, extraction, scoring, or query rewriting, whereas the launch communication speaks of a general-purpose model. The API is OpenAI-compatible, and streaming is active by default.