MiniMax M3 has a pretty serious parameter stack—wait 10 days for the open-source release and you’ll see what everyone means by “so worth it.”

CryptoWorld News reports that MiniMax officially released its large model M3 today. M3 is currently the only open-source large model that brings together three cutting-edge elements—programming, ultra-long context, and native multimodality—and it plans to formally open-source its weights within 10 days. It reaches international leading levels in code generation, agents, and desktop control, and can be experienced in MiniMax code, token plan, and API. M3 pioneered a sparse attention architecture called MSA, which aggregates hit queries across KV blocks, making memory access 4 times faster than Flash-sparse-attention. With a 1 million context window, the new architecture reduces per-token computation to one twentieth of the previous generation, achieving 9x prefill and 15x decoding speedups. On SWE-bench pro, M3 scored 59.0%, surpassing GPT-5.5 and Gemini 3.1.