MiniMax Just Launched M3.1 Flash: 428B MoE Agentic Model at $0.15/M Input

Why it matters

MiniMax has launched MiniMax M3.1 Flash, a multimodal mixture-of-experts model available through its API, and confirmed that the stealth model Space Bunny was an earlier build of the same system. The model matches MiniMax M3 in scale, with a large total parameter count and a much smaller active portion, and it accepts text, image, and video input with a very long context window. Reasoning is always on and cannot be turned off, with several effort levels available to developers. According to MiniMax's own benchmarks, the model performs strongly on agentic office tasks, leading Automation-Bench, while coding results are mixed and it trails Opus 5 on Terminal Bench. All scores are vendor-reported with no independent replication yet. Model weights are scheduled for release in late October, and the license has not been disclosed. The launch positions MiniMax to compete on cost per task for agentic workloads.