Celeris-1's Diffusion Bet on Speed

Celeris.ai’s launch announcement for celeris-1: diffusion-based inference instead of autoregressive decoding, a 157ms p50 response latency (about 15x faster than GPT-5-mini, 17x faster than GPT-5), and a 76% score on MMLU-Pro against 78% for GPT-5-mini and 81% for GPT-5.

On raw throughput, celeris-1 reaches 1,280 tokens per second against 144 for Gemini 3.5 Flash-Light, measured on a reconstructed version of Artificial Analysis’s benchmark dataset. Replies on X pushed back quickly on the “near-GPT-5” framing, given that gap on the same benchmark.

Fichamento

#

• Celeris-1 replaces autoregressive generation with a diffusion-based inference architecture, which the company presents as the source of its speed gains without losing ground on the MMLU-Pro score cited alongside it.
• Its p50 response latency is 157ms, about 15x faster than GPT-5-mini and 17x faster than GPT-5.
• On MMLU-Pro it scores 76%, against 78% for GPT-5-mini and 81% for GPT-5.
• Throughput reaches 1,280 tokens per second, compared with 144 for Gemini 3.5 Flash-Light, on a reconstructed version of Artificial Analysis’s benchmark dataset.
• The company frames its mission as maximizing useful intelligence delivered per unit of time, and points to celeris.ai/blog for build details and the full benchmark set.
• Replies on X questioned the “near-GPT-5” label given the actual MMLU-Pro gap, and several commenters asked whether the model would ship as open weights.