Celeris Launches Ultra-Fast Diffusion-Based Language Model Celeris-1
Celeris, an AI research lab based in San Francisco, has launched its flagship language model, Celeris-1. Built on a novel diffusion architecture rather than the traditional autoregressive method used by most large language models (LLMs), Celeris-1 delivers near-GPT-5 intelligence with significantly faster response times—up to 24 times quicker in some benchmarks. The new model generates responses through parallel processing, starting with an approximate version of a full output and refining it rapidly. This approach enables real-time performance suitable for interactive applications such as voice agents, live translation, and autonomous systems requiring low-latency reasoning. Celeris-1 achieves 75.9% accuracy on the MMLU-Pro benchmark compared to Mercury 2's 63.7%, while generating text at a rate of 1,664 tokens per second. Celeris claims this breakthrough opens new possibilities for AI integration in latency-sensitive environments like agentic tool orchestration and real-time data extraction workflows. The model is now available via an OpenAI-compatible API.
Topics: AI Celeris-1 diffusion models LLM