MiniMax M3 Review: MSA Architecture, 1M-Token Context, Native Multimodal — Generational Leap or Benchmark Theater?

MiniMax M3 (June 1, 2026) abandons the M2 MoE design for a completely new Sparse Attention architecture, adds native video/image input, and claims SWE-Bench Pro parity with GPT-5.5 at 5-10% of the cost. We review the architecture, benchmarks, pricing, open-weight status, and what the M2.7 license history means for M3.