MiniMax open‑sources M2.7
MiniMax announced it has open‑sourced its M2.7 model and says it achieved top scores on SWE‑Pro (56.22%) and Terminal Bench 2 (57.0%), with the model now available on Hugging Face. The release and benchmark claims were posted on the company's social channel and point to community access to the weights and code base (x.com/MiniMax_AI/status/2043132047397659000).
MiniMax has released M2.7 as open weights on Hugging Face, putting one of its latest coding-focused models into public download. (huggingface.co) The model card says M2.7 is a 229 billion-parameter text-generation model under a modified MIT license, and MiniMax also published a public GitHub repository for the release. (huggingface.co) (github.com) MiniMax says M2.7 scored 56.22% on SWE-Pro and 57.0% on Terminal Bench 2, two tests built to measure whether an artificial intelligence system can handle longer, messier software-engineering work instead of short code snippets. (minimax.io) (scaleapi.github.io) (tbench.ai) SWE-Pro is a benchmark of 1,865 software problems drawn from 41 repositories, with tasks that can require patches across multiple files and hours or days of work by a human engineer. Terminal Bench 2 measures whether an agent can complete jobs inside a sandboxed command-line environment, including system administration and security tasks. (scaleapi.github.io) (tbench.ai) That makes this release part of a wider shift in open-model competition: companies are no longer selling only chat quality, but also whether a model can act like a software worker that reads logs, edits code, runs tools, and recovers from failure. MiniMax’s own model card says M2.7 is aimed at “complex agent harnesses,” productivity tasks, and multi-agent collaboration. (huggingface.co) MiniMax says M2.7 can use “Agent Teams,” a setup where multiple specialized agents split work the way a small engineering group might divide debugging, planning, and execution. The company also says an internal version of the model improved a programming scaffold over more than 100 rounds and raised performance by 30%. (huggingface.co) (github.com) The company is not new to open releases. In November 2025, MiniMax said it was open-sourcing MiniMax M2, another model it described as built for agents and code. (minimax.io) MiniMax describes itself as a global foundation-model company founded in early 2022, and its investor-relations site says its model lineup now spans text, audio, image, video, and music systems. In March 2026, the company reported 2025 revenue of $79.0 million, with more than 70% from international markets. (minimax.io) (ir.minimax.io) (minimax.io) The open question is whether outside developers reproduce MiniMax’s benchmark numbers with the same scaffolds and settings. The weights are public now, which means the next test shifts from a company post to community runs. (huggingface.co) (scaleapi.github.io)
