MiniMax M3.1 Flash
Points and comments are a snapshot, not live.
MiniMax claims M3.1 Flash matches top models on agentic and coding benchmarks at 70% lower cost.
MiniMax launched M3.1 Flash, a next-generation model claiming a generational leap in agentic tasks, coding, computer use, and cybersecurity. It scores 73.7% on DeepSWE v1.1, 76.2 on BankerToolBench, 61.7 on OSWorld V2, and 82% on CyberGym. The model is 3x faster at inference with median speeds of 170+ TPS. Pricing is $0.15 per 1M input tokens, $0.60 per 1M output, and $0.015 per 1M cached input, 70% cheaper than its predecessor. It is available via API, M Plan, and MiniMax Code; weights to follow.
What commenters are saying
Commenters compare M3.1 Flash to DeepSeek V4.1 Flash. Some Reddit comments say it is not as good; one commenter found it weaker on DeepSWE and declared it dead. Others note the model has 428B total parameters with 23B active, similar to DeepSeek 4.1 Flash. The thread questions whether 'flash' is a misnomer for large models. There is skepticism about benchmark relevance and wall-clock speed. Multiple duplicate submissions were flagged, and one user notes the minimax.io domain may be shadowbanned. The Pac-Man demo ran but took 30 minutes.