Celeris Labs, an AI research lab that describes its mission as building the world’s fastest language models, announced celeris-1 on X on 27 July 2026. The company says the general-purpose model delivers near-GPT-5-level intelligence with responses 15 times faster than comparable frontier systems, built on a new inference architecture that applies diffusion techniques rather than the sequential decoding most large language models use.
The specific numbers are notable. Celeris Labs reports a p50 response latency of 157 milliseconds and throughput of 1,280 tokens per second, both well outside the range typical of current frontier chatbots. Neither figure comes with an independent benchmark run, a third-party leaderboard placement, or a published comparison methodology. The company says a companion post detailing how celeris-1 was built, including full benchmarks, is available. At publication, no outside lab has replicated the GPT-5 comparison or verified the throughput figure under load, so the entire performance claim rests on Celeris Labs’ own measurements, announced through a single post on X rather than a peer-reviewed paper or a public leaderboard submission.
Diffusion language models differ from autoregressive ones like GPT-5 in a basic way. Autoregressive decoding generates text one token at a time, each token conditioned on everything generated before it, so latency tends to scale with output length. Diffusion models instead start from a noisy or masked draft of the entire output and refine it in parallel across a fixed number of denoising steps, the same idea that powers image generators like Stable Diffusion, applied to text. That parallelism decouples latency from output length, which is the plausible engineering reason behind a flat 157 millisecond figure rather than one that grows with every additional token.
The approach is not unique to Celeris Labs. Diffusion-based text models have appeared from other startups over the past two years, built on the argument that the architecture suits fast, parallel inference better than transformer decoding. What sets celeris-1 apart, on paper, is the intelligence claim layered on top of the speed claim. Celeris Labs has not disclosed model size, training compute, or the benchmark suite behind “near-GPT-5-level,” so that phrase describes the company’s characterization of its own model, not a verified score.
For teams building latency-sensitive products (voice agents, real-time coding assistants, anything where token-by-token wait time shows up as a user-facing metric), a genuine 15x speedup at comparable quality is worth testing directly rather than taking on faith. Until celeris-1 is available for outside evaluation, treat the 157ms figure as an opening claim from the company that made it, not a settled industry benchmark.
Celeris Labs announced celeris-1 on X on 27 July 2026.
