MiniMax open M2.7 model

MiniMax has released M2.7 as an open model aimed at software engineering and other agent-style tasks. (minimax.io) Large language models predict the next token, but “agent” systems wrap them with tools, memory, and step-by-step actions so they can search, edit code, and run commands. MiniMax said M2.7 was built for those longer workflows rather than single-turn chat. (minimax.io) MiniMax said M2.7 scored 56.22% on SWE-Bench Pro, a test of realistic software bugs and engineering tasks, and 57.0% on Terminal Bench 2, which measures how well models handle command-line work. The company also reported 55.6% on VIBE-Pro and 39.8% on NL2Repo. (github.com) SWE-Bench Pro is designed around enterprise-style engineering problems, with 1,865 problems drawn from 41 actively maintained repositories. Its maintainers say model runs are allowed up to 250 turns, which makes it a test of sustained problem-solving rather than short code snippets. (scaleapi.github.io) Terminal Bench 2 is a separate leaderboard for terminal-based agents, and its public rankings are much higher than 57.0% for the top entries. As of April 2026, the first-ranked system on that board was listed at 82.9%, which means MiniMax’s 57.0% is a solid result but not the top public score on that benchmark. (tbench.ai) MiniMax describes M2.7 as a sparse mixture-of-experts model, which is a design that activates only part of the network for each request instead of using every parameter every time. Fireworks AI lists the model at 228 billion parameters with a 196,600-token context window, plus serverless and dedicated deployment options. (fireworks.ai) MiniMax said the model took part in its own training loop by updating memory, building skills for reinforcement-learning experiments, and refining parts of its learning process based on results. The company’s GitHub repository says an internal version improved a programming scaffold over more than 100 rounds and raised performance by 30%. (minimax.io) (github.com) The release is already being packaged for production systems. NVIDIA’s NIM container catalog lists MiniMax-M2.7 as available on April 13, 2026, with an OpenAI-compatible application programming interface, SGLang serving, and FP8 quantization on supported NVIDIA graphics processing units. (catalog.ngc.nvidia.com) That puts M2.7 into the part of the market where companies want open weights they can run themselves for coding, search, office automation, and tool use. MiniMax’s pitch is that M2.7 is not just a chatbot with code skills, but a model meant to sit inside longer-running software workflows. (catalog.ngc.nvidia.com) (minimax.io)