MiniMax has debuted M3.1 Flash, a 428B-parameter mixture-of-experts model designed for agentic workloads. Only about 23B parameters are active for any given token, which helps keep inference costs down. The model also supports a 1M-token context window, making it suitable for long-document retrieval, multi-step tool use, and other agent-heavy tasks.
Pricing is a notable part of the launch: $0.15 per million input tokens and $0.60 per million output tokens, putting it in a competitive range for high-volume applications. The open release extends further, with model weights scheduled to arrive on October 23, allowing teams to self-host or fine-tune the model.
Since this report covers only a single announcement, there are no conflicting claims to reconcile. The main open question is how M3.1 Flash will perform in real-world agentic workflows once developers start benchmarking it against other large MoE models.