MiniMax Just Open-Sourced a 428B Coding Model — Can You Actually Run It?

MiniMax has recently released an open weights 428 billion parameter coding and agent model, but the reality of running it is far more complex than the headline suggests. The model, named M3.1 Flash, was made available on the API on October 19 and its weights on October 23 via Hugging Face and ModelScope. The model boasts an impressive 428 billion total parameters, is designed for coding and agent tasks, and features a 1 million token context window.

It is capable of processing native text, image, and video input, making it well-suited for computer-use agents. The model's architecture is sparse, with only 23 billion active parameters per token, and it employs full-repo agents and long-horizon computer-use tasks. The FP8 checkpoint is approximately 220 gigabytes in size. The model's performance is quantified through benchmarks, including 73.7% on DeepSWE v1.1, 61.7 on OSWorld 2.0, and 82.2 on CyberGym.

Community support for the model is robust, with vLLM, SGLang, Docker Model Runner, llama.cpp, Ollama, and LM Studio all providing support. In terms of hardware requirements, 4x NVIDIA DGX Spark are needed to run the model, with each box costing around $4,000. This makes the self-hosted minimum cost around $16,000. The MiniMax Community License permits commercial use for businesses with annual revenue under $20 million, requiring only a notification email and a credit display.

For larger enterprises, written authorization is required. This license is more generous for indie hackers and early-stage startups, allowing for commercial use with attribution.