MiniMax M3: Open-weight model with a million-token context challenges proprietary leaders

Details

Chinese AI company MiniMax has released M3, positioning it as the first open-weight model to combine a one-million-token context window, native multimodal capabilities, and strong coding performance in a single system. The key technical enabler is 'MiniMax Sparse Attention,' which processes only relevant data blocks rather than the full context, cutting compute to one-twentieth and speeding up input processing by more than nine times compared to standard attention. Benchmark results, self-reported by MiniMax: 59% on SWE-Bench Pro, placing M3 ahead of GPT-5.5 and Gemini 3.1 Pro but just behind Opus 4.7. On BrowseComp (autonomous web search), M3 scores 83.5, ahead of Opus 4.7's 79.3. The company ran internal 12-hour autonomous research experiments — M3 independently reproduced an ICLR 2025 paper over 18 commits and 23 figures, and optimized a CUDA matrix multiplication kernel for Nvidia Hopper GPUs without human intervention. The model is currently available via API, with weights to be published shortly. The sparse attention architecture represents a genuine technical contribution if the efficiency claims hold — enabling a million-token context without proportional compute cost has been an unsolved problem for open models. Benchmarks are self-reported and require independent validation. M3 was trained using a simulator framework that mimics real developer workflows — multi-turn requirement refinement, intermediate result reactions, cross-context task continuation — which the company credits for strong long-horizon autonomy.