MiniMax M3 Open-Weight Model Beats GPT-5.5 with 1M Token Context

Chinese AI startup MiniMax has released M3, a large language model it claims is the first open-weight system to combine frontier coding performance, a 1-million-token context window, and native multimodal capabilities. On SWE-Bench Pro, M3 scored 59%, surpassing OpenAI's GPT-5.5 and Google's Gemini 3.1 Pro, though trailing Anthropic's Opus 4.7. The model also scored 83.5 on BrowseComp, ahead of Opus 4.7's 79.3. M3 is built on a new MiniMax Sparse Attention (MSA) architecture that reduces per-token compute for 1M-token contexts to just 1/20th of its predecessor, with input processing over 9x faster and response generation over 15x faster. The model was trained on roughly 100 trillion tokens with interleaved text and images from the start, enabling native multimodal understanding without a bolt-on vision layer. All benchmark results are company-reported and have not been independently verified.

M3 is available immediately via MiniMax's API platform, with token plan subscriptions starting at $20/month (~1.7B tokens) up to $120/month (~9.8B tokens). Model weights and a full technical report are scheduled for release on Hugging Face and GitHub within 10 days, though the license terms remain unspecified—'open-weight' may carry usage restrictions. MiniMax also plans to open-source its in-house coding agent, MiniMax Code, built on M3. The release comes amid competition from other Chinese AI labs like DeepSeek and Zhipu AI. If validated, M3 could democratize access to frontier-level AI coding and long-context reasoning, challenging the dominance of proprietary models from Western labs.