MiniMax releases M3 frontier model for coding agents with one-million-token context
The AMW Read
M3 is a new frontier model from an existing Chinese lab, meaningfully updating the player map in segment 01; open-weight strategy and coding-agent focus exemplify §5 recurring patterns.
MiniMax releases M3 frontier model for coding agents with one-million-token context
MiniMax, the Shanghai-based AI lab listed on the Hong Kong Stock Exchange, has launched M3, a frontier model designed for coding agents with a one-million-token context window and native multimodal input. The company claims M3 scores 59.0% on SWE-Bench Pro and 66.0% on Terminal-Bench 2.1, outperforming GPT-5.5 and Gemini 3.1 Pro on SWE-Bench Pro while approaching Claude Opus 4.7. MiniMax says it will open-source the model weights within 10 days and has also released its MiniMax Code agent product alongside the model.
Why it matters: M3 represents a direct challenge from a Chinese AI lab in the coding-agent segment, where Anthropic's Claude Code, OpenAI's coding agents, and Google's Gemini tools currently dominate. The one-million-token context window paired with MiniMax's Sparse Attention architecture, which the company claims cuts per-token compute by 20x at that context length, addresses the cost problem that has limited long-context agent adoption in developer workflows. The open-weight promise also puts MiniMax in a rare position: a publicly listed Chinese AI company trying to balance open-source distribution with monetization via hosted APIs and agent subscriptions, a tension that has defined the open-weight strategy debate.
Expert take: The independent benchmark results will determine whether M3 becomes part of the daily developer stack. MiniMax's claims — especially the Sparse Attention efficiency gains — need third-party validation before enterprise procurement teams can rely on them. What is already clear is that M3 forces the market to take a Chinese coding competitor seriously at a moment when coding benchmarks are shifting toward original long-horizon tasks rather than short fixes. The next 10 days, when weights are promised to drop, will test whether MiniMax can execute on its open-source commitments and whether M3 holds up outside the company's own infrastructure.
