AI · 2h ago
MiniMax M3: Open-Weight Model with 1M-Token Context and Sparse Attention
MiniMax released M3, the first open-weight model combining 59% SWE-Bench Pro score, 1M-token context, and native multimodal input. Its MiniMax Sparse Attention reduces compute by 28.4x at 1M tokens by attending only to 2,048 relevant tokens per query. API pricing at $0.30/M input tokens undercuts rivals by 10-20x.
Meridian48 take
The sparse attention breakthrough is real, but self-reported benchmarks and restrictive licensing temper the open-weight promise.
minimax-m3sparse-attention