Celeris launches diffusion LLM API for sub-second agent loops

Why it matters

Celeris-1 gives diffusion language models a second commercial supplier and tests whether lower latency, rather than another capability gain, can reshape voice and agent software.

Tom Hamer and Jesse Clark launched Celeris-1 on July 27th, giving developers access to a diffusion-based language model that Celeris says can complete short reasoning calls in milliseconds and generate long responses at 1,664 tokens per second.

The San Francisco AI lab made Celeris-1 available through an OpenAI-compatible API, according to an announcement distributed through PR Newswire. Celeris is positioning the model for voice agents, classification, extraction, tool selection and other workloads where a delay at each model call compounds across a larger software process.

"Speed changes what you can build," Hamer, Celeris's co-founder and CEO, said in the announcement. The line captures the founders' bet: model vendors have spent years pushing benchmark performance higher, while latency remains an architectural constraint for applications that need to reason during a live conversation or run dozens of internal calls before returning an answer.

Hamer and Clark previously co-founded Marqo, an AI product-discovery and search infrastructure provider. Hamer worked as an engineer at AWS and earned a machine learning master's from Cambridge. Clark, Celeris's chief technology officer, led machine learning work at Amazon Robotics, worked as a principal scientist at Stitch Fix and held physics research posts at Stanford and University College London. On his personal site, Clark says he led development of perception models for Amazon's Robin warehouse robot and designed a continual-learning framework used in Amazon Robotics AI.

Celeris lists Lightspeed Venture Partners, Blackbird Ventures and January Capital as backers. The same investor group backed Marqo, giving Hamer and Clark institutional support from their previous AI-infrastructure company as they move from retrieval systems into model architecture.

Parallel generation instead of a token-by-token queue

Most widely deployed language models generate from left to right, predicting one token and feeding it back into the model before producing the next. That sequential process limits how much generation can be parallelized, even when inference providers use faster chips or optimized serving software.

Celeris-1 instead refines multiple parts of a response during each generation step. Celeris compares the process to bringing a blurry image into focus, an analogy borrowed from diffusion image models. A technical post published July 23rd gives a more precise description of Celeris's research direction: hybrid architectures that combine sequential generation where language requires it with parallel decoding elsewhere.

That distinction matters. Celeris-1 is presented as a diffusion model, while Celeris's research thesis retains some sequential computation to preserve coherence. The commercial advantage will depend on how well the architecture identifies which portions of a response can be generated together without sacrificing instruction following or reasoning accuracy.

Celeris is entering a category opened by Inception's Mercury models, which also use diffusion to generate tokens in parallel. Inception describes Mercury 2 as a reasoning diffusion model capable of more than 1,000 tokens per second on standard Nvidia GPUs. The emergence of a second commercial API gives developers another architecture to test against conventional autoregressive models and specialized inference providers such as Groq and Cerebras.

The speed claims come from Celeris's own tests

Celeris reports that Celeris-1 scored 75.9% on MMLU-Pro with a median server response time of 158 milliseconds. In the same Celeris-run benchmark, GPT-5 scored 81.9% at 2.046 seconds and GPT-5 mini scored 78.5% at 2.495 seconds. Celeris disabled reasoning budgets for the compared models and used each provider's fastest available configuration.

The "near-GPT-5 intelligence" description in Celeris's marketing rests on that single MMLU-Pro comparison. It does not establish comparable performance across coding, long-context work, tool use or multi-step production agents. The test also reports model or server processing time for most providers, while Gemini results were measured end to end from a colocated client.

Celeris's methodology introduces additional qualifications. Mercury 2 was moved to a different, zero-shot forced chain-of-thought setup after struggling with the standard answer format, and Celeris substituted a previously published score for Gemini 2.5 Flash after its own run produced formatting failures. The methodology note also cites a 63.1% Mercury result while the main table reports 63.7%.

The throughput comparison is a separate test. In a July 22nd output-speed report, Celeris said Celeris-1 generated a median 1,664 tokens per second and completed roughly 1,000-token outputs in 0.58 seconds. Mercury 2 generated 324 tokens per second in that run, while GPT-5 generated 69. Celeris sent 50 requests per model using ten long-form prompts, with outputs capped at 1,024 tokens.

Those figures have not been independently reproduced. They are still relevant as an engineering claim because Celeris publishes the workload, timing basis and output lengths instead of presenting an unexplained tokens-per-second figure. Developers considering the API will need to measure latency from their own regions and with their own prompt distributions, concurrency levels and tool-calling patterns.

A deliberately narrow first product

The initial Celeris API is aimed at compact tasks. Celeris's developer documentation recommends Celeris-1 for short, structured responses and limits the prompt plus maximum output allocation to 8,192 tokens. Maximum output settings must be positive multiples of 256.

That makes Celeris-1 a targeted piece of inference infrastructure. Classification, extraction, query rewriting and agent routing rarely require enormous context windows, but each call can sit directly in a latency-sensitive path. Shaving seconds from those steps can have a larger product effect than accelerating a single chatbot response.

Celeris charges $2 per million input tokens and $6 per million output tokens. Celeris also offers custom enterprise terms, dedicated clusters and VPC deployment. Existing OpenAI clients can connect by changing the base URL and API key, reducing the integration work required to run an evaluation.

The API design reveals the founders' go-to-market strategy. Hamer and Clark are starting with workloads where diffusion's speed advantage can be measured immediately and where a constrained context window is less damaging. Voice agents and multi-step software agents give Celeris a practical test of its core claim: intelligence becomes a different infrastructure product when each reasoning call fits inside the response time of an ordinary interface.