Skip to main content
Radar

Five or fewer confirmed AI changes, with original sources, on mornings when something changed.A free source-linked morning brief.

Start free brief
BenchLM

Celeris

Data verified
Celeris Labs is an AI research company building low-latency language models. Its first model, Celeris-1, uses diffusion-based decoding and is served through an OpenAI-compatible API.
Model count
1
Best model
N/A
Celeris-1
Avg. top 3
N/A
Latest release
July 2026
Public-ranked
0
Supported
0

Provider profile

What is Celeris?

Celeris Labs is an artificial intelligence research company focused on reducing language-model inference time. Tom Hamer and Jesse Clark founded the company after previously founding Marqo. Celeris says its research spans model training, diffusion-based architecture, and the production systems required to serve language models at interactive speeds.

Celeris-1 is the company’s first public model. It is a proprietary, general-purpose diffusion language model released on July 22, 2026. The hosted API accepts text and images, returns text, supports streaming, and exposes a 131,072-token context window. Current public access is limited to the United States.

Celeris at a glance

Company
Celeris Labs
Founders
Tom Hamer and Jesse Clark
First model
Celeris-1
Architecture
Diffusion language model
API
OpenAI-compatible chat completions
Public availability
United States

What the published evidence supports

The source-attached ledger now includes four Celeris-published results: 75.9% on MMLU-Pro, 93.7% on GSM8K, 80.8% on IFEval, and 81.4% on DROP. Celeris ran the MMLU-Pro test set with five-shot chain-of-thought examples, strict answer-format scoring, and the reasoning budget set to zero. The results are provider-published rather than independently reproduced, so they remain visible with their sources while the model stays outside the overall ranking.

Celeris separately reports 158 ms p50 response time on the MMLU-Pro run and 1,664 output tokens per second at p50 on a 1,000-token workload. The independent runtime series carries a separate current snapshot. The workloads differ, so the results stay separate.

Where Celeris-1 fits

The model is aimed at short, structured, latency-sensitive work: routing, extraction, classification, query rewriting, agent steps, and interfaces that cannot hide a long generation behind a loading state. OpenAI compatibility lowers the integration cost because existing SDK calls can point at Celeris with a base-URL change.

The 131,072-token ceiling covers the prompt plus requested output, and the official model guide still recommends another model for very long-form generation. One provider-run knowledge benchmark also cannot establish coding, agentic, multilingual, instruction-following, or visual-reasoning performance. Test the actual request shape before putting it on a critical path.

Celeris-1 model profileInspect the exact benchmark row, price, context, and source ledger.Compare LLM API pricingCompare Celeris-1 token rates with other hosted language models.Read the speed methodologySee why runtime figures need one comparable measurement pipeline.

Provider models

Browse all models
ContextPricePublic score
Celeris-1Current128K$0.20 / $0.7019320.61s—

Celeris FAQ

What is Celeris AI?

Celeris Labs is an AI research company focused on low-latency language-model training, architecture, and inference. Its first public model is Celeris-1, a proprietary diffusion language model served through an OpenAI-compatible API. The company was founded by Tom Hamer and Jesse Clark, who previously founded Marqo.

What model does Celeris offer?

Celeris currently lists Celeris-1 as its public model. It accepts text and images, returns text, and has a 131,072-token context window shared by the prompt and requested output. The provider recommends another model for very long-form generation, and public API availability is currently limited to the United States.

Is Celeris-1 a diffusion LLM?

Yes. Celeris describes Celeris-1 as a diffusion language model that can decode multiple tokens per model invocation instead of relying only on sequential, token-by-token generation. The company has not released the model weights or a complete architecture specification, so that description remains a provider claim rather than an independently inspected implementation.

How much does the Celeris-1 API cost?

Celeris lists pay-as-you-go pricing at $0.20 per million input tokens and $0.70 per million output tokens, metered separately. The plan includes streaming and the OpenAI-compatible API. Enterprise customers can ask for volume pricing, dedicated clusters, custom limits, VPC deployment, and a service-level agreement.

How fast is Celeris-1?

Celeris reports 1,664 output tokens per second at p50 on its long-form workload and 158 ms p50 response time on its MMLU-Pro run. The model page also carries a separate current cross-provider runtime snapshot. The workloads differ, so the results should not be treated as replications.

Primary sources

Radar

Celeris release history

Full release history

Compare Celeris models