so there's no flash<br>of the wrong palette before React hydrates. Mirrors the useTheme hook<br>(storage key "rw-theme"; classes rw-theme-dark|light|ascii). -->

Celeris-1 tops AI speed ranking at 2,158 tokens per second - RuntimeWire

RuntimeWire

You're browsing RuntimeWire with JavaScript disabled. Articles and<br>navigation work fully. Interactive features — search, comments,<br>and newsletter signup — require JavaScript.

Why it matters

Independent speed leadership gives diffusion language models a concrete commercial proof point. Celeris still has to show that the advantage survives production workloads without widening the intelligence gap.

Celeris-1 has taken the top speed slot among comparable nonreasoning models in Artificial Analysis' current benchmark, generating 2,157.9 output tokens per second. The result gives founders Tom Hamer and Jesse Clark an independent data point for their bet that diffusion can make language models fast enough to sit inside voice interfaces, coding tools and multi-step agent loops.

Celeris said in an August 4th thread on X that an earlier benchmark snapshot measured Celeris-1 at 2,038 output tokens per second, twice the rate of the next model, and ranked it first across a 591-model comparison for output speed, response time, time to first token and time per intelligence task. Artificial Analysis' current public summary reports an even higher throughput figure and labels Celeris-1 first for speed among 80 models in its comparable price and capability class. It records a time to first token of 0.64 seconds.

The independent result sharpens the pitch Hamer and Clark made when they publicly launched Celeris-1 in late July. The pair previously co-founded Marqo, which builds semantic and visual search infrastructure for online retailers. Celeris says Hamer earned a machine-learning master's degree from Cambridge and built infrastructure at AWS. Clark led machine learning at Amazon Robotics, worked as principal machine-learning scientist at Stitch Fix and held physics research posts at Stanford and University College London.

Their shift from search infrastructure to model inference follows the same underlying thesis: latency constrains which AI systems developers can put into production. Celeris lists Lightspeed Venture Partners, Blackbird Ventures and January Capital as backers.

Diffusion moves generation into parallel

Most commercial language models use autoregressive decoding, generating each token after the previous token. Celeris describes a different process: Celeris-1 begins with a rough version of a response and refines the sequence over several parallel passes. The method adapts diffusion, an architecture widely associated with image generation, to text.

At Celeris-1's July 27th launch, Celeris described the model as the second commercially available diffusion language-model API after Inception's Mercury. Clark said the research question was whether a new architecture could preserve strong reasoning while operating within real-time latency limits.

Celeris' own benchmark measured 1,664 output tokens per second, below the subsequent Artificial Analysis results. Benchmark throughput can change with prompt length, endpoint load and test configuration, so the figures are snapshots rather than fixed properties of the model.

In Celeris' MMLU-Pro test, Celeris-1 scored 75.9% with a median server-reported response time of 158 milliseconds. GPT-5 scored 81.9% at 2,046 milliseconds, while Mercury 2 scored 63.7% at 257 milliseconds. Celeris used zero reasoning budgets for Celeris-1, GPT-5 and the other reasoning-capable models in the comparison, and tested Mercury 2 in instant mode. Most latency figures were reported by the model providers; Gemini endpoints were measured end to end from a colocated client.

Artificial Analysis offers a more restrained reading of Celeris-1's intelligence. Its current page gives the model a score of 12 on the Artificial Analysis Intelligence Index, ranking it 40th among 80 comparable models. The benchmark therefore supports Celeris' speed case more strongly than its claim of near-frontier intelligence.

That profile defines the immediate market. Celeris' developer documentation points builders toward short, repeated tasks such as classification, extraction, scoring, query rewriting, live translation and agent tool calls. Shaving seconds from each request can compound across an agent that makes dozens of model calls before returning an answer.

Celeris-1 is delivered through an OpenAI-compatible API, allowing developers to switch endpoints without rewriting the basic request format. Pricing is $0.20 per million input tokens and $0.70 per million output tokens. Enterprise plans include dedicated clusters, custom rate limits and virtual private cloud deployment.

For Hamer and Clark, the speed ranking arrives less than two weeks after Celeris-1's July 24th release. That timing turns an architecture argument into a developer-acquisition tool: builders...