MiniMax Launches Agent-Native AI Models
AI company MiniMax has launched its M2.5 and M2.1 models, positioning them as built for “real-world productivity” and complex, multi-language programming. The company's associated Agent platform emphasizes a focus on managing requirements and intent over simple code generation, arguing that “code is cheap.”
Shanghai-based MiniMax, founded in 2021 by former SenseTime researchers, has rapidly emerged as a key player in the AI space, recently listing on the Hong Kong Stock Exchange in January 2026. The company is backed by major tech investors, including Alibaba and Tencent, and reached a valuation of $4 billion after a $300 million funding round in July 2025. The new M2.5 model is an open-weight Mixture-of-Experts (MoE) architecture with 230 billion total parameters, but only 10 billion are active during any given task. This design aims for high performance on complex coding and agentic workflows while maintaining efficiency, allowing it to run on a handful of GPUs. M2.5 demonstrates strong performance on benchmarks like SWE-Bench Verified, rivaling models like Claude Opus 4.6 in speed but at a fraction of the cost. MiniMax's Agent platform is designed for more than just code generation; it focuses on executing multi-step tasks autonomously. The platform can plan solutions, use external tools like web search and code interpreters, and adjust its approach based on the results. This "interleaved thinking" allows the model to tackle complex workflows from start to finish, such as building landing pages or conducting detailed research. The company has a "full-stack" strategy, developing its own foundation models while also building consumer-facing applications. Their AI companion app, Talkie, has already attracted over 212 million users globally, providing a massive dataset that helps improve their models. This dual approach of proprietary models and large-scale consumer products distinguishes MiniMax from many of its competitors.
