Celeris-1
Celeris · 2026-07-27
Celeris-1 is Celeris's debut diffusion-based language model, announced July 27, 2026. It generates and refines whole responses in parallel rather than token-by-token, targeting real-time and short structured agentic calls (classification, routing, extraction, judging) at roughly 1,664 tokens per second with sub-200ms median response times, and is served exclusively through the Celeris API. It accepts text and images, returns text or structured tool calls, and supports opt-in reasoning through chat_template_kwargs.enable_thinking.