On June 1, MiniMax released the M3 model. In a nutshell: an open-weight model that packs a 1-million-token context, cutting-edge coding capabilities, and native multimodality. MiniMax says it is the first to combine all three.
The real work is in 'sparse attention'
M3 is built on a new architecture called MSA (MiniMax Sparse Attention).
To understand what it solves, you need to know the old problem: standard full attention requires compute that scales quadratically with context length—so long contexts become prohibitively expensive. That's why 'million-token context' has been talked about for so long but remains costly to use.
MSA targets this directly. MiniMax's numbers: at 1 million tokens, M3 uses only one-twentieth the compute per token of its predecessor M2; prefilling is over 9x faster, and decoding over 15x faster.
In other words, million-token context is now not just 'usable' but 'affordable'.
Benchmarks: self-reported, impressive
Coding is M3's main focus. MiniMax released these scores:
- SWE-Bench Pro: 59.0% (surpasses GPT-5.5, Gemini 3.1 Pro, close to Opus 4.7)
- Terminal-Bench 2.1: 66.0%
- MCP Atlas: 74.2%
- OSWorld (computer control): 70.06%
If these numbers hold, they are quite competitive for an open-weight model.
But a dose of caution: these are all self-reported by MiniMax and have not been independently verified. Some overseas media have already run headlines like 'Frontier Claims, Unverified Benchmarks.' The smart move is to wait for the weights and technical report—expected in about ten days—and let the community test them.
Pricing also announced
MiniMax also laid out its plans:
- Plus: ~1.7 billion tokens per month, $20
- Max: ~5.1 billion tokens per month, $50
- Ultra: ~9.8 billion tokens per month, $120
The API is already available. Model weights and the technical report are expected on Hugging Face and GitHub within about ten days.
In the broader context, this is another push by Chinese open-source LLMs toward 'long context + capable agents.' DeepSeek, Qwen, and Kimi have been competing in this area for a year. MiniMax's differentiation lies in architectural efficiency—not who has more parameters, but who can compute faster and cheaper at the million-token scale.
This cuts at the moat of proprietary models: 'long context is too expensive.'
Sources: CocoLoop, MiniMax Releases MiniMax M3 with MSA Architecture Supporting 1M-Token Context (MarkTechPost); MiniMax M3: Open-weight model with a million-token context challenges proprietary leaders (The Decoder); MiniMax M3 Open-Weight Coding Model: Frontier Claims, Unverified Benchmarks (TechTimes)