Celeris-1

Summary

Celeris product launch of celeris-1, a closed general-purpose LM using diffusion-style generation instead of token-by-token AR. Site: MMLU-Pro 75.9% at p50 response 158 ms vs GPT-5 81.9%/2.0s and GPT-5 mini 78.5%/2.5s (reasoning budget 0 / Mercury 2 instant). Output p50 1,664 tok/s on ~1k-token long-form vs GPT-5 69. Tweet: p50 157 ms (~15× GPT-5-mini, ~17× GPT-5), MMLU-Pro 76% vs 78/81, and 1,280 tok/s vs Gemini 3.5 Flash Lite 144 on a reconstructed Artificial Analysis TPS set. OpenAI-compatible API at inference.celeris.ai. No paper or weights.

Keywords

celeris · diffusion-lm · latency · mmlu-pro · x

Topics

diffusion LMs, low-latency inference

Research notes

• Primary: product page. Discord/X https://x.com/Celeris_ai/status/2080442996403933630 via fxtwitter. Closed API; access left blank rather than guess paid. Benchmarks are vendor-reported. Not a hosted corpus, so no datasets_local row.