Perceive Beating AI News Flash: Celeris has released a hybrid diffusion Agent model, Celeris-1 Magnus, based on Qwen3.8-27B, which they claim can achieve over 500 Token/s in many scenarios.

In 97 banking customer service tasks for τ³-Banking, Magnus achieved a completion rate of 41.2% after enabling low reasoning, with a median time of 55 seconds. In the same setting, GPT-5.6 Sol scored 38.1% in completion rate and 79 seconds.

While traditional models usually generate one Token at a time, Magnus can generate multiple Tokens at once and then iteratively correct them. For content that needs to be strictly generated in order, it is processed sequentially. Therefore, each computation can advance more Tokens forward.