MiniMax M3 launches with 1M context

MiniMax M3 Frontier represents a major victory for efficient long-context architectures, showing that proprietary sparse attention mechanisms can achieve frontier performance with vastly reduced hardware requirements.

* Proprietary MSA architecture reportedly slashes prefill and decoding costs to as low as 1/20th of previous generations, presenting a massive threat to compute-heavy rivals.

* A native 1 million token context window combined with robust multimodal capabilities makes it a prime candidate for complex, long-horizon agentic and coding workflows.

* Strong benchmark numbers on SWE-Bench Pro show that open-weight/cost-efficient models are rapidly narrowing the gap with closed frontier giants.