Celeris, an artificial intelligence research lab focused on advancing frontier intelligence through building ultra-fast LLMs, announced Celeris-1, the lab's flagship language model. Built from the ground up on a novel diffusion architecture, Celeris-1 breaks away from the decade-long standard of autoregressive token generation, establishing a new foundation for interactive, low-latency AI systems.

"Speed changes what you can build," said Tom Hamer, Co-founder and CEO of Celeris. "We optimize for intelligence delivered per unit of time, and Celeris-1 is the first demonstration of that principle."

The Paradigm Shift: From Sequential Tokens to Parallel Diffusion

To date, state-of-the-art Large Language Models (LLMs) have relied almost exclusively on autoregressive decoding. Under this paradigm, generation is strictly sequential: the model predicts one token at a time, using every previously generated token as context to predict the next.

Diffusion models take a fundamentally different approach. Diffusion is the same technique behind modern image generation, applied to language: rather than producing an answer word by word, the model starts with a rough version of the entire response and improves it over a few rapid passes, the way a blurry image comes into focus. Rather than each word being generated one by one, the entire sequence is generated and refined together, which dramatically speeds up response times and throughput.

"Our objective wasn't simply to make language models faster. It was to explore whether new architectures could preserve frontier-level reasoning while operating within real-time latency constraints," said Jesse Clark, Co-founder and CTO of Celeris. "Celeris-1 is our first public demonstration that this direction is possible, and we believe there is significantly more headroom ahead."

Benchmark Results

Celeris-1 is the second commercially available diffusion language model API, following Inception's Mercury. It generates 1,664 tokens per second, over 5 times the output speed of Mercury 2, previously the fastest model in its class, at substantially higher reasoning accuracy: 75.9% versus 63.7% on MMLU-Pro. Against frontier baselines like GPT-5, Celeris-1 generates output at 24 times the speed.

What faster intelligence unlocks

By reducing the latency barrier between model capability and human interaction, Celeris-1 makes frontier-level reasoning viable in settings where responsiveness is essential. It is engineered for structured, latency-sensitive production workloads, including agentic tool orchestration, classification, data extraction, and real-time voice pipelines:

• Interactive latency: At a median of 158 milliseconds, responses land below the threshold at which humans perceive delay, making AI viable in live conversational audio and immediate control loops where traditional model delay causes drop-offs.
• Fast agentic loops: Autonomous agents routinely make dozens of internal reasoning calls for routing, validation, and tool selection. When each call completes in milliseconds instead of seconds, those savings compound with every step.
• Real-time interfaces: Voice agents, live translation, and in-session assistants all require models that reason and respond within the timescale of human workflows.

Availability & Access
Celeris-1 is now available through Celeris' OpenAI-compatible API. Developers can integrate Celeris-1 into existing workflows by updating their API base URL and key. Sign up and get an API key at celeris.ai.