Chinese AI lab MiniMax has unveiled its latest model, the M2.7, which reportedly matches leading closed models in performance benchmarks. The model achieved a score of 56.22% on the SWE-Pro benchmark, designed for software engineering tasks, and 57.0% on Terminal Bench 2. It also attained an ELO score of 1495 on the GDPval-AA benchmark, reflecting its capabilities in real-world knowledge work across various occupations. These scores place the M2.7 close to models such as Claude Opus 4.6, Sonnet 4.6, and GPT-5.4, positioning it as a competitive player in the AI landscape.
The M2.7 is structured as a 230 billion parameter Mixture of Experts model, deploying only 10 billion parameters during each inference pass. MiniMax claims that this architecture allows for high-quality output without the substantial computational costs typical of top-tier models. Notably, the M2.7 underwent an unprecedented self-optimization process without human intervention, reportedly improving performance by 30% through autonomous rounds of refinement.
However, the excitement surrounding the M2.7 has been clouded by a significant licensing change. Shortly after releasing the model’s weights, MiniMax revised the terms to restrict commercial use, necessitating written authorization for such applications. While non-commercial use remains free and unrestricted—allowing for research, personal projects, and fine-tuning—this new requirement poses challenges for developers seeking to incorporate the model into commercial products or hosted services.
This licensing shift has ignited fierce discussions across platforms like Hacker News and Hugging Face, where developers expressed confusion over MiniMax labeling the revised license as “MIT-style.” The original MIT license allows for commercial use by default, while the new “Modified-MIT” label creates ambiguity, according to critics. Ryan Lee, MiniMax’s Head of Developer Relations, provided insight into the rationale behind the change. He explained that some hosting providers were deploying subpar versions of earlier MiniMax models, leading to reputational damage for the company and a poor user experience.
Lee articulated the company’s frustration, stating, “They walk away thinking MiniMax is mid. We get the reputational bill, the user gets a bad experience, and the serious hosting providers who do the work properly get drowned out in the noise.” He pointed out that the previous fully permissive license left MiniMax unable to address these issues effectively. “If the license has edge cases that hurt legitimate community use, tell us. We’d rather fix the text than defend it,” Lee added.
This marks a notable shift for MiniMax, which had built its reputation on open releases, including the M2 under the MIT license in October 2025 and the M2.5 under the same terms in February 2026. The M2.7 is the first model to deviate from this trend, coming just months after the company went public on the Hong Kong Stock Exchange, raising approximately $620 million with backing from prominent investors including Alibaba and Abu Dhabi’s sovereign wealth fund.
The move aligns with a broader trend in the Chinese tech landscape, where companies traditionally seen as champions of open-source development are increasingly experimenting with proprietary models. Reports indicate that Alibaba’s Qwen team has shifted towards proprietary development, following leadership changes, while Xiaomi has also released its latest MiMo v2 models under a closed-source license. The narrative that Chinese labs focus on open-source AI while their U.S. counterparts lean toward closed models is evolving.
For developers interested in utilizing the M2.7 for commercial purposes, Lee assures that the authorization process will be “fast and reasonable,” although the shift in licensing has raised concerns in the community. As the landscape of AI development continues to change, the future implications of such licensing decisions may significantly impact developers and companies looking to leverage advanced AI technologies.
See also
Google Commits $10M to Train 40,000 Workers in AI for Manufacturing Jobs
Stanford AI Report Reveals U.S.-China Model Gap Closing, 95% of AI Projects Fail
Lumentum Secures $2B Nvidia Investment, Boosting AI Orders Through 2028
Google DeepMind Hires Philosopher Henry Shevlin to Explore Machine Consciousness and Ethics
Microsoft’s AI Integration Fuels 20% Revenue Growth Amid Intensifying Competition
