MiniMax M3.1 Flash 正式发布并开源
MiniMax M3.1 Flash 正式上线并开源，主打 agentic coding 与 computer-use，官方称推理快近 3 倍、agent 任务成本降 70%。
Remember Space Bunny? It was an early look at MiniMax M3.1 Flash, which is now officially launched. M3.1 Flash is open source, giving developers the opportunity to explore and build. The model is making a strong case in agentic coding and computer-use benchmarks: → 73.7% on DeepSWE v1.1 for complex software engineering → 76.2 on BankerToolBench for multi-step financial workflows → 61.7 on OSWorld V2 for computer-use tasks → 82% on CyberGym for cybersecurity I also tested its coding capabilities by building NEON VELOCITY, a 3D cyberpunk racing game. The idea was to create an interactive racing experience with futuristic city visuals, 3D environments, and gameplay mechanics. What stood out during the build: → Handling a complex coding task beyond a basic website → Bringing visual design and interactive elements together → The importance of token efficiency when refining and improving a project through multiple iterations Another interesting part is the cost-effectiveness. MiniMax reports nearly 3× faster inference and 70% lower usage costs on agentic tasks, making it an interesting option for developers working on projects that require repeated coding and refinement. The key isn't just benchmark scores. It's how performance, token efficiency, speed, and cost come together in real workflows.