Celeris releases Magnus hybrid diffusion model for low-latency agent tasks
Celeris has released Celeris-1 Magnus, a hybrid diffusion model based on Qwen3.8-27B, designed for tool loops and long tasks, and supporting 131,000 token contexts.
Celeris releases Celeris-1 Magnus, a hybrid diffusion language model optimized for agent workloads.The company stated that the model is based on Qwen3.8-27B and supports text and tool calls, 131,000 token contexts, and inference modes that can be switched on request.
Magnus is now available via an OpenAI-compatible API, with an input price of $0.20 per million tokens and an output price of $0.70. Celeris reported a completion rate of 41.2% and a median task time of 55 seconds on τ³-bench banking; these benchmark figures are from company evaluations and are subject to independent verification.
refer to:Celeris product page