So, MiniMax quietly released [M2.5](https://www.minimax.io/models/text) on February 12th, and honestly, the numbers are kind of absurd. We’re talking about an open-weight model that scores 80.2% on SWE-Bench Verified — just 0.6 points behind Claude Opus 4.6. Let that sink in for a second. An open model is now essentially neck-and-neck with the best proprietary coding model on the planet.

The architecture is a 230B parameter Mixture of Experts setup, but it only activates 10B parameters per forward pass. That’s how they keep it fast and cheap. And when I say cheap, I mean comically cheap. Running M2.5 continuously at 100 tokens per second for a full hour costs about one dollar. Compare that to Claude Opus, and you’re looking at roughly 1/20th the price. MiniMax is basically saying “frontier intelligence, but make it affordable.”

There are two flavors: the standard [M2.5](https://huggingface.co/MiniMaxAI/MiniMax-M2.5) and a Lightning version that cranks out 100 tokens per second. Both are available on [GitHub](https://github.com/MiniMax-AI/MiniMax-M2.5) and [HuggingFace](https://huggingface.co/MiniMaxAI/MiniMax-M2.5) under a modified MIT license, so you can actually self-host and fine-tune this thing. The model was trained using their Forge RL framework across hundreds of thousands of real-world environments, which helps explain why it doesn’t just benchmark well — it actually handles messy, practical coding tasks.

The buzz has been real. [VentureBeat](https://venturebeat.com/technology/minimaxs-new-open-m2-5-and-m2-5-lightning-near-state-of-the-art-while) covered the cost angle extensively, the [South China Morning Post](https://www.scmp.com/tech/article/3343395/chinas-minimax-releases-cheap-ai-model-designed-real-world-productivity) noted MiniMax’s Hong Kong-listed shares jumped 15.7% on the news, and [The Information](https://www.theinformation.com/briefings/chinas-minimax-launches-new-flagship-ai-model-m2-5) highlighted its strong coding and agentic capabilities. Over on [Hacker News](https://news.ycombinator.com/item?id=46990403), the thread blew up with developers pointing out that M2.5 is the first open model to genuinely surpass Claude Sonnet on coding benchmarks. That’s a milestone worth paying attention to.

Here’s what makes this actually matter beyond the hype: MiniMax claims that 80% of their own newly committed code is now generated by M2.5, and 30% of their internal tasks across departments are handled autonomously by the model. Whether those internal numbers hold up externally remains to be seen, but the SWE-Bench score isn’t something you can hand-wave away. If you’ve been waiting for an open model that can genuinely compete with the top-tier closed options for real coding work, M2.5 is the strongest candidate we’ve seen yet.