Celeris, a new AI research lab backed by Lightspeed Venture Partners, has released Celeris-1, a language model built on diffusion architecture rather than the token-by-token approach almost every major LLM uses today. The company says it hits a median response latency of 158 milliseconds while scoring 75.9% on MMLU-Pro, and claims to generate output 24 times faster than GPT-5.
That’s a bold set of numbers from a vendor announcement, so it’s worth stepping back and looking at what diffusion architecture actually changes, and why speed has become such a pointed argument in the current AI market.
Why Sequential Generation Became the Bottleneck
Since GPT-3, nearly every frontier model has relied on autoregressive decoding: the model predicts one token, adds it to the context, then predicts the next. It’s a proven approach, but it’s inherently sequential. Every token has to wait for the one before it.
Diffusion models take a different path, one borrowed from image generation tools like Stable Diffusion. Instead of writing left to right, the model starts with a rough draft of the full response and refines it in parallel across a handful of passes. Researchers at Cornell, describing a related architecture called STAR-LDM, frame this as adding a planning phase that works on a semantic level before committing to specific words, rather than making token-by-token decisions with no ability to revise earlier choices.
This isn’t a brand-new idea. Inception Labs, a Stanford spinout, has been shipping commercial diffusion models under the Mercury name since last year. Its most recent release, Mercury 2, reaches over 1,000 tokens per second on Blackwell GPUs and is pitched specifically at agentic workloads, where a single task can trigger dozens of model calls in sequence. Inception describes the shift plainly: less typewriter, more editor revising a full draft at once.
Does the Quality Hold Up?
The obvious question with any faster architecture is what gets traded away. Academic research so far suggests less than expected. A study out of Qualcomm AI Research found diffusion models build more globally distributed representations across layers than autoregressive models, with enough redundancy that researchers could skip nearly 19% of compute while retaining over 90% of performance on math and coding benchmarks. Dream 7B, a diffusion model built on top of Qwen2.5, reported reaching parity with its autoregressive counterpart on general language tasks while showing specific strength in planning-heavy scenarios. A separate empirical study of code generation found diffusion models solving certain tasks that autoregressive models still struggle with.
That said, independent, apples-to-apples benchmarking of diffusion models against top-tier autoregressive systems is still thin. Most of the strongest comparisons come from research labs measuring against open models like LLaMA and Qwen, not against the current frontier of closed commercial systems. Celeris’s specific claims against GPT-5 haven’t been independently verified yet.
Where the Speed Argument Actually Matters
The use case getting the most attention is agentic AI. Mitch Ashley, VP and practice lead for software lifecycle engineering and AI-native software engineering at The Futurum Group, puts the math in blunt terms: “Latency compounds in agent pipelines. Twenty chained calls at three seconds each is a minute of dead time that no operations team will ship. The same twenty calls at 300 milliseconds is a workflow people will actually use.”
Inception has made a similar argument with Mercury 2, and voice AI vendors have echoed it: at sub-200ms response times, a model starts operating within the same timescale as a live conversation.
But Ashley cautions against treating speed as a stand-in for governance. “Faster inference does not settle what agents may do. It only removes the excuse that automation is too slow. Set latency budgets per agent step now, while pipeline designs are still cheap to change.”
Whether Celeris-1 lives up to its own numbers in production is a separate question from whether diffusion architecture deserves attention. The research base for the latter question is growing, and it’s no longer just Inception making the case alone. For enterprise teams building agent pipelines where every millisecond compounds, this is a category worth tracking, even if the leaderboard is still unsettled.