MiniMax open-sources MiniMax-M2 for coding and agentic workflows

MiniMax released MiniMax-M2, a 230-billion-parameter mixture-of-experts model with 10 billion active parameters, under the permissive MIT licence. The company built it specifically for coding and agentic workflows — multi-file edits, coding-run-fix loops and test-validated repairs, plus tool-use tasks involving shells, browsers and code execution — rather than as a general-purpose chat model, and said its compact active-parameter count kept feedback loops fast in agent settings.

MiniMax priced API access at roughly 8% of the cost of Claude Sonnet 4.5 while claiming close to double the inference speed, and reported that the model’s composite score on the Artificial Analysis benchmark suite ranked first among open-weight models globally. On individual benchmarks the comparison to Sonnet 4.5 was mixed: MiniMax reported stronger results on BrowseComp, a web-research benchmark, but a lower score on SWE-bench Verified, a measure of real-world software-engineering task completion.

The release continued a pattern through 2025 of Chinese labs — DeepSeek, Alibaba’s Qwen, Moonshot’s Kimi and others — using open weights and steep price undercutting to compete with US closed frontier labs on coding and agentic tasks specifically, an area where enterprise demand and willingness to pay were both concentrated. MiniMax offered the API free for a limited period after launch to encourage adoption before settling into its stated per-token pricing, a promotional tactic other open-weight labs had also used to build developer share quickly. MiniMax followed M2 with an M2.1 update in December 2025.