Home
MiniMax Open-Sources M3 Large Model with MSA Architecture, 1M Context, Rivaling Overseas Flagships
MiniMax Xiyu Technology unveiled its next-generation large model, the MiniMax M3, on June 1, 2026. It stands as China’s first open-source model to combine elite-level programming, an ultra-long 1M context window, and native multimodal capabilities, designed to rival leading overseas closed-source models across the board.

To address context expansion bottlenecks in complex agent tasks, M3 introduces a sparse attention architecture (MSA) at the foundation level. Compared to conventional approaches, it delivers more precise KV partitioning and operator-level optimizations, achieving four times the computational speed of comparable open-source solutions. With a 1M context length, the per-token computational cost drops to just one-tenth of the previous generation. It accelerates the prefill stage by over 9x and the decoding stage by 15x.

Through native trillion-scale interleaved data training, M3’s semantic space achieves deep integration. It outperforms GPT-5.5 and Gemini 3.1 Pro in authoritative software engineering and multimodal evaluations like SWE-Bench Pro. In extreme task tests, M3 demonstrates robust long-thread autonomous planning: it independently reproduced an ICLR top paper’s experiments within 12 hours, ran 24 hours straight without reference code while calling tools nearly 2,000 times, boosted FP8 matrix multiplication hardware utilization on Hopper architecture from 7.6% to 71.3%, and autonomously orchestrated the full “data-training-iteration” pipeline on the open PostTrainBench.
Related article
Stripe to Acquire OpenRouter as AI Model Routing Gains Momentum
Stripe has agreed to acquire OpenRouter, an AI model-routing platform that gives developers access to hundreds of models through a single interface. The deal adds model selection and routing to Stripe’s existing work around AI usage and token-based b
Former OpenAI Researcher Barrett Zoph Returns to Google to Support Gemini Development
Barret Zoph, who previously abruptly left and joined OpenAI due to a disagreement with Thinking Machines Lab CEO Miriam Mularati, has once again changed employers. Alphabet's Google announced that Zoph will return as Vice President of Research, bring
How Recursive Intelligence Raised $335M at $4B Valuation in 4 Months
The co-founders of startup Ricursive Intelligence seemed destined to be co-founders.Anna Goldie, CEO, and Azalia Mirhoseini, CTO, are so well-known in the AI community that they were among those AI engineers who “got those weird emails from Zuckerber
Related Special Topic Recommendations
Comments (0)
0/500

To address context expansion bottlenecks in complex agent tasks, M3 introduces a sparse attention architecture (MSA) at the foundation level. Compared to conventional approaches, it delivers more precise KV partitioning and operator-level optimizations, achieving four times the computational speed of comparable open-source solutions. With a 1M context length, the per-token computational cost drops to just one-tenth of the previous generation. It accelerates the prefill stage by over 9x and the decoding stage by 15x.

Through native trillion-scale interleaved data training, M3’s semantic space achieves deep integration. It outperforms GPT-5.5 and Gemini 3.1 Pro in authoritative software engineering and multimodal evaluations like SWE-Bench Pro. In extreme task tests, M3 demonstrates robust long-thread autonomous planning: it independently reproduced an ICLR top paper’s experiments within 12 hours, ran 24 hours straight without reference code while calling tools nearly 2,000 times, boosted FP8 matrix multiplication hardware utilization on Hopper architecture from 7.6% to 71.3%, and autonomously orchestrated the full “data-training-iteration” pipeline on the open PostTrainBench.
Stripe to Acquire OpenRouter as AI Model Routing Gains Momentum
Stripe has agreed to acquire OpenRouter, an AI model-routing platform that gives developers access to hundreds of models through a single interface. The deal adds model selection and routing to Stripe’s existing work around AI usage and token-based b
Former OpenAI Researcher Barrett Zoph Returns to Google to Support Gemini Development
Barret Zoph, who previously abruptly left and joined OpenAI due to a disagreement with Thinking Machines Lab CEO Miriam Mularati, has once again changed employers. Alphabet's Google announced that Zoph will return as Vice President of Research, bring
How Recursive Intelligence Raised $335M at $4B Valuation in 4 Months
The co-founders of startup Ricursive Intelligence seemed destined to be co-founders.Anna Goldie, CEO, and Azalia Mirhoseini, CTO, are so well-known in the AI community that they were among those AI engineers who “got those weird emails from Zuckerber











