MiniMax M3 hits production with 1M-token sparse attention

Block-sparse attention reduces N² scaling to make 1M context windows feasible; 9x prefill, 15x decode speedup over dense attention at the cost of reimplementing attention kernels and multimodal preprocessing pipelines.

June 3, 2026

Summary

Long-context inference (codebases, documents, agentic loops) becomes cost-competitive. Developers targeting production agentic systems can now evaluate a model built for tool-use at scale without prohibitive latency or KV-cache overhead.

Why it matters

Long-context inference (codebases, documents, agentic loops) becomes cost-competitive. Developers targeting production agentic systems can now evaluate a model built for tool-use at scale without prohibitive latency or KV-cache overhead.

Implementation verdict

Replaces dense attention implementations and KV-cache management strategies. Requires custom kernel work (block-major reordering, sparse paged attention integration, decode scoring optimization) and gateway-level multimodal preprocessing. Ready now via Together AI endpoint; self-hosted deployment demands kernel engineering expertise.

Sources

  1. 1.1M-token context window, native multimodality, and an architecture that demands serious engineering to serve efficiently
  2. 2.brings a speed up of more than 9x in the prefilling stage and more than 15x in the decoding stage
  3. 3.The attention computation itself no longer scales as N^2 with context length, thus making it very suitable for long context workload
  4. 4.MSA significantly lowers the wall time percent of the actual attention computation per iteration

Dev Signal

Get briefs like this in your inbox — free, 3x a week.

100+ sources compressed into one 4-minute read. Ranked, cited, implementation-ready.