Celeris-1 ranks #197 of 227 on the public leaderboard with a score of 28.36/100. It does not yet have enough sourced coverage for a verified position.
Celeris-1 is a proprietary model with a 131,072 tokens context window. No explicit reasoning mode is documented in this profile.
Available through Celeris's OpenAI-compatible API in the United States. The model accepts text and image input, returns text, and is intended for short, structured, latency-sensitive tasks; Celeris recommends another model for very long-form generation.
Official exact-value snapshot from Celeris's July 2026 benchmark report and model documentation. Celeris reports 75.9% on the full MMLU-Pro test set with five-shot chain-of-thought exemplars, strict answer-format scoring, and the reasoning budget set to zero. The provider also reports 158 ms p50 response time and 1,664 output tokens per second at p50 in separate first-party measurements; those speed figures remain outside BenchLM's runtime field because the runtime pipeline uses a single independent methodology.
4 of 726 tracked benchmark slots currently have displayable evidence. Missing categories stay blank.
Its strongest eligible category is Knowledge at #163. particularly effective for knowledge-intensive tasks like research, analysis, and factual Q&A.
Model research
What Celeris-1 is built for
Celeris-1 is a proprietary diffusion language model released on July 22, 2026. Celeris built it for short requests that sit inside an interactive path, including structured extraction, routing, classification, query rewriting, and repeated agent steps. The OpenAI-compatible API accepts text and images, returns text, and streams responses by default.
Diffusion is the architectural distinction. Autoregressive models normally commit to one token after another; Celeris says its system combines parallel decoding with sequential processing where language needs it. The model weights and full training specification are not public, so the profile records that description without implying an independent architecture audit.
How to read the speed claims
Celeris reports 1,664 output tokens per second at p50 on 49 successful timed requests from a 50-request workload. It also reports a 158 ms p50 response time alongside the 75.9% MMLU-Pro result. Both tests publish prompt shape, sample count, timing basis, and important limitations.
The cross-provider runtime snapshot supplies the independent speed card. It does not reproduce Celeris’s 1,664-token result because the prompts and timing methods differ. Each number belongs to its own workload.
Price, context, and the practical limit
Pay-as-you-go pricing is $0.20 per million input tokens and $0.70 per million output tokens. The context window is 131,072 tokens shared by the prompt and requested output, API access is currently limited to the United States, and Celeris recommends using another model for very long-form generation.
That makes Celeris-1 a focused candidate rather than a default model for every workload. Published evidence now covers knowledge, grade-school math, instruction following, and passage comprehension, but not coding, tool use, multilingual tasks, or long-context retrieval. A production decision still needs a replay set built from the requests the application will actually send.