Post

MiniMax M3.1 Flash: Redefining Agent Performance and Cost

MiniMax says its new M3.1 Flash model is faster and cheaper for agentic workloads, but Hacker News users strongly questioned whether its benchmark scores match real-world quality.

The company reports up to 200+ output tokens/s, 70% lower agent-task costs than its predecessor, and results including 73.7% on DeepSWE and 61.7 on OSWorld V2. It is available through MiniMax’s API and products; the company says model weights will come later. These are vendor-reported figures, not independent evaluations.

The 214-comment HN discussion was notably skeptical: several users said the earlier anonymous “Space Bunny” version disappointed in practice and argued benchmark scores are unreliable, while another pointed to a niche coding benchmark where it performed well. The thread’s clearest takeaway is that independent task experience matters more than headline benchmark claims.