← Back to feed

Two labs using diffusion instead of autoregressive decoding now hold the top two speed slots on Artificial Analysis. Celeris-1, launched July 27, and Inception’s Mercury 2 together represent the fastest models available through a public API. Neither uses custom inference silicon.

The numbers on Celeris-1 are specific. In end-to-end measurements from AWS us-east-1, it generated a median of 1,664 tokens per second and returned complete 1,000-token responses in 0.58 seconds. Against the providers that have invested most heavily in custom hardware, the outcome is blunt: Cerebras runs its wafer-scale engine at 1,480 tokens per second — Celeris reaches 1,596, 8% faster, from a standard EC2 instance. Groq’s custom LPU delivers 460 tokens per second, 3.5x off the pace.

The accuracy trade-off is visible but bounded. On MMLU-Pro with reasoning budget set to zero, Celeris-1 scores 75.9% in 158 milliseconds. Gemini 3.5 Flash Lite reaches 83.0% but takes 1.23 seconds — about 8x longer. GPT-5 scores 81.9% in 2.05 seconds. For tasks where benchmark accuracy matters more than speed, Celeris-1 is not the answer. For every intermediate step in a pipeline, the trade-off runs the other way.

Agent Pipelines Are the Argument

The compound math in agent loops is where the speed delta becomes decisive. A 10-step pipeline generating 1,000-token outputs per step completes in roughly six seconds with Celeris-1. The same pipeline takes 31 seconds with Mercury 2, 76 seconds with Gemini Flash, and 148 seconds with GPT-5. Sub-second per-step latency lets teams run techniques that normally belong in offline processing — sampling multiple candidate responses and keeping the best, regenerating a document instead of patching it — within interactive latency budgets.

Celeris-1 published one specific claim about the two distributions not overlapping: the slowest Celeris-1 request across 49 timed runs was 814 tokens per second; the fastest single request from any other model was 662. The fastest and slowest ends of the distributions do not overlap.

Why Commodity Hardware Wins Here

The Celeris-1 vs Groq/Cerebras comparison is unusual because it inverts the typical expectation. Groq’s LPU and Cerebras’ wafer-scale chip exist specifically to accelerate inference. Celeris runs on standard NVIDIA GPUs in AWS us-east-1. The speed advantage is architectural, not silicon.

Diffusion language models generate multiple tokens per forward pass through parallel refinement rather than left-to-right sequential decoding. Each forward pass refines the entire output span simultaneously. Longer outputs do not compound delay the way they do in autoregressive models: a 300-token response and a 1,000-token response follow a similar time curve rather than scaling linearly with length. That property is what makes the throughput numbers hold under longer outputs.

Celeris-1 is the second commercially available diffusion language model API. Mercury 2 from Inception launched in February 2026 at approximately 1,000 tokens per second on NVIDIA Blackwell GPUs. Celeris-1 claims 5.1x Mercury 2’s speed at the median while also scoring higher on MMLU-Pro: 75.9% versus Mercury 2’s 63.7%.

Leaderboard Snapshot

Artificial Analysis currently ranks Celeris-1 and Mercury 2 at the top of its output speed table, followed by Gemini 2.5 Flash-Lite and Ling 3.0 Flash. For TTFT latency, Gemini 2.5 Flash-Lite and Command A+ lead the table. The intelligence column, which weights quality across a broader benchmark suite, still places Claude Opus 5 and Claude Fable 5 at the top.

Celeris is backed by Lightspeed Venture Partners and describes its research objective as treating latency as “a fundamental research problem rather than an engineering optimization.” Celeris-1 is available through an OpenAI-compatible API.