← Back to feed

Celeris Labs has launched Celeris-1, the second commercially available diffusion language model. It generates 1,664 tokens per second at p50 and returns complete responses to 1,000-token prompts in 0.58 seconds end-to-end — more than five times faster than Inception’s Mercury 2, the previous speed leader, and roughly 24 times faster than GPT-5.

The benchmark quality result, 75.9% on MMLU-Pro, is not frontier intelligence. Gemini 3.5 Flash Lite scores 83.0%, GPT-5 mini 78.5%. The claim Celeris is making is different: Celeris-1 is the model that fits inside the interactive latency budget.

Speed by the Numbers

Model	Tokens/sec (p50)	Response time p50	MMLU-Pro
Celeris-1	1,664	0.58 s	75.9%
Mercury 2 (Inception)	324	3.10 s	63.7%
Gemini 3.5 Flash Lite	224	4.55 s	83.0%
GPT-5 mini	99	10.31 s	78.5%
GPT-5	69	14.78 s	81.9%

The gap between 158ms and the nearest autoregressive model is not marginal. Diffusion models produce many tokens per model invocation rather than one token per forward pass. Celeris is the second lab after Inception to bring this architecture to a public commercial API.

Both latency figures — Celeris-1’s 0.58 seconds and Mercury 2’s 3.10 seconds — are measured end-to-end from AWS us-east-1. The autoregressive models use server-reported processing time, which excludes network overhead. The comparison runs against Celeris-1: its figures include real network latency and it still leads by 5x.

The Context Constraint

Celeris-1 has an 8,192-token context window. The default output is 2,048 tokens, with the maximum bounded by whatever fits in context. The lab explicitly recommends using a different model for long-form generation. Pricing is $2 per million input tokens and $6 per million output tokens. The API is OpenAI-compatible. US access only at launch.

These constraints are deliberate, not oversights. The targets Celeris is explicit about: classification, extraction, scoring, routing, query rewriting, and the repeated structured steps that fill out most production agent pipelines. Those tasks are typically short-context. They run thousands of times per day in real systems.

A 10-step agent pipeline that spends 0.58 seconds per generation completes its compute work in about 6 seconds with Celeris-1. The same pipeline on GPT-5 spends roughly 148 seconds. That difference compounds across every concurrent user session.

At sub-second response times, techniques that are usually reserved for offline processing fit inside interactive latency budgets. Five full resamples of a 1,000-token response take about 3 seconds — close enough to real-time that regenerating a result becomes a plausible user interaction rather than a background task.

The Diffusion Speed Tier

Two diffusion models now have commercial APIs: Mercury 2 (Inception) and Celeris-1. Together they define a speed tier that sits roughly an order of magnitude below frontier models on latency while operating in an overlapping intelligence range — above GPT-4-class, below GPT-5.

Mercury 2 runs at 324 tokens/second and 63.7% MMLU-Pro. Celeris-1 runs at 1,664 tokens/second and 75.9% MMLU-Pro. The new entrant is ahead of Mercury 2 on both speed and quality. Neither model is competitive with GPT-5 or Gemini 3.5 Flash on intelligence for complex tasks, but both clear the threshold for tasks where speed is the binding constraint.

The speed tier is not a niche. Most production inference requests are not hard reasoning problems. Classification, intent detection, slot filling, and document routing are what most enterprises actually run at scale. For those workloads, waiting 14 seconds for a GPT-5 response is not a feature — it is a cost and latency problem that diffusion architecture addresses.

Key Numbers

• 1,664 tokens/sec (p50), 1,906 tokens/sec (p90)
• 158ms p50 response time on MMLU-Pro workload (server-reported)
• 75.9% MMLU-Pro accuracy (reasoning budget disabled)
• $2/M input, $6/M output
• 8,192-token context window
• Available now via Celeris API, US only