MiniMax-M2 Goes Open Source, Surpasses Claude Opus 4.1 on New AI Intelligence Benchmark

MiniMax has open-sourced its new flagship AI model, MiniMax-M2, positioning it as one of the most efficient AI systems for coding and agent-based tasks available today.

Designed as an “agent- and code-native” model, MiniMax-M2 is built specifically for end-to-end developer workflows and agent reasoning.

Despite having a total of 230 billion parameters, the model activates only 10 billion at a time, delivering near-state-of-the-art performance in a more compact and cost-effective form.

MiniMax claims that MiniMax-M2 delivers results at roughly 8% of the cost of Claude Sonnet and runs nearly twice as fast. According to the AI Analysis Intelligence Index v3.0, MiniMax-M2 scored 61 points, ranking eighth overall—outperforming Anthropic’s Claude Opus 4.1, which scored 59.

The AI Analysis benchmark aggregates results from 10 key evaluations, including MMLU-Pro, GPQA Diamond, AIME 2025, SciCode, and Terminal-Bench Hard, to assess general reasoning and tool-use capabilities.

MiniMax-M2 now ranks among the strongest open-source models, surpassing Qwen 3 72B (58 points) and DeepSeek-V3.2 (57 points). While not the absolute top open-weight model, it leads the pack in this particular benchmark.

Benchmark comparisons highlight its highly competitive coding performance: it scored 46.3 on Terminal-Bench, beating both Claude Sonnet 4.5 and Gemini 2.5 Pro, and achieved a 44 on BrowseComp—significantly higher than Claude Sonnet 4.5’s 19.6.

MiniMax is offering limited-time free access to MiniMax-M2 through its Agent and API platform and has released the model weights on Hugging Face and GitHub for local deployment.

With benchmark results that place it ahead of Claude Opus 4.1, MiniMax-M2 underscores the growing strength of open-source AI models engineered to balance affordability, speed, and advanced reasoning for real-world coding and agent applications.