AI BriefingMiniMax-M3 Debuts, Surpassing GPT-5.5 and Gemini 3.1 Pro on Key Benchmarks at Just 5-10% of the Cost
Key point
China's MiniMax has released M3, a multimodal model that captures both performance and cost-efficiency through its innovative MSA architecture.
Details
Chinese AI startup MiniMax has officially launched 'M3', a multimodal large language model (LLM) featuring frontier-level coding performance and a 1 million token context window. M3 will be distributed under an Open Weights license via Hugging Face and GitHub within the next 10 days, enabling local optimization for enterprises.
Core Technology and Economics At the heart of M3 is the 'MiniMax Sparse Attention (MSA)' architecture, which solves the computational cost problem of conventional transformers. This technology reduces computational requirements by 1/20 when processing 1 million tokens and boosts decoding speed by 15x. Building on this, MiniMax achieved a groundbreaking price of $1.50 total cost per 1 million tokens, which is only 5-20% of the cost of major U.S. models.
Key Performance Metrics
- SWE-Bench Pro: Scored 59.0%, demonstrating coding capabilities that surpass GPT-5.5 and Gemini 3.1 Pro.
- BrowseComp: Scored 83.5%, exceeding existing models in web browsing and information retrieval performance.
MiniMax also unveiled the 'MiniMax Code' agent product line, which is powered by M3 and autonomously writes and modifies code.
This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.
Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.
Related
Strategy for Owning the AI Intelligence Layer: A 4-Step Roadmap to Building In-House AI Based on Open-Weight Models