MiniMax M2.7 open‑sourced
The MiniMax M2.7 model was open‑sourced and reported top scores on developer benchmarks like SWE‑Pro and Terminal Bench 2, with NVIDIA publicly congratulating the team. (x.com) NVIDIA also highlighted GPU‑accelerated endpoints and developer tooling such as NemoClaw in posts linked to the release. (x.com)
A large language model is software that predicts the next token, one chunk at a time; MiniMax has now released M2.7 for developers to download and run themselves. (minimax.io) MiniMax published M2.7 on April 11, 2026 through its site, GitHub, Hugging Face, and ModelScope, alongside an API and NVIDIA-hosted endpoints. (minimax.io) (github.com) (build.nvidia.com) (huggingface.co) (modelscope.cn) The model is a sparse mixture-of-experts system, which works like a large panel where only a few specialists answer each request instead of the whole room speaking at once. NVIDIA’s model card lists 230 billion total parameters, 10 billion active parameters, 256 experts, and a 204,800-token context window. (developer.nvidia.com) (build.nvidia.com) MiniMax says M2.7 scored 56.22% on SWE-Pro, 57.0% on Terminal Bench 2, 55.6% on VIBE-Pro, 76.5 on SWE Multilingual, 52.7 on Multi SWE Bench, and 39.8 on NL2Repo. NVIDIA repeated those benchmark numbers in its April 11 post about the release. (github.com) (developer.nvidia.com) Those tests are aimed at developer work that looks more like a real terminal or repository than a coding puzzle. MiniMax says the model is built for log analysis, bug troubleshooting, refactoring, code security, machine-learning workflows, and office document editing. (github.com) (minimax.io) (build.nvidia.com) MiniMax also tied the release to “Agent Teams,” its term for multiple model roles working together on one task. The company says M2.7 can use dynamic tool search and maintain stable role identity across multi-agent runs. (github.com) Part of the pitch is that the model helped tune its own training loop. MiniMax wrote that an internal M2.7 variant optimized a programming scaffold over more than 100 rounds and improved performance by 30%. (github.com) NVIDIA used the launch to promote its own stack around the model. Its April 11 blog said M2.7 is available through NVIDIA, can run with NemoClaw and OpenShell for always-on assistants, and has inference optimizations in vLLM and SGLang for mixture-of-experts workloads. (developer.nvidia.com) (github.com) The release also comes with licensing caveats. NVIDIA’s model card says use is governed by the NVIDIA Open Model License and lists “Modified MIT License” for MiniMax M2.7, while MiniMax’s earlier M2 releases on Hugging Face were published under MIT. (build.nvidia.com) (huggingface.co 1) (huggingface.co 2) By Monday, April 13, the immediate result was clear: MiniMax had turned a benchmark-heavy model launch into a broad distribution push across open repositories, cloud endpoints, and NVIDIA tooling. (github.com) (build.nvidia.com) (developer.nvidia.com)
