MiniMax M3 百万token上下文以5%算力成本击败GPT-5.5
minimax发布了m3，这是一个拥有100万token上下文和稀疏注意力机制的开放权重模型，可将计算量降低95%。它在编码基准测试中击败了gpt-5.5，这标志着随着开放模型匹敌专有能力，专有护城河正在消退。
chinese ai company minimax launched m3, an open-weight model combining native multimodality, a 1-million-token context, and coding performance competitive with proprietary leaders like gpt-5.5 and opus 4.7. the model uses 'minimax sparse attention' to cut compute costs by 95% and boost processing speeds, allowing it to achieve 83.5 points on browsecomp and 59% on swe-bench pro. internal tests showed m3 performing autonomous tasks such as reproducing research papers and optimizing nvidia gpu kernels for 24 hours without intervention. this development signals a significant shift in the ai landscape, as open models begin to rival proprietary systems in complex reasoning and long-context tasks.