MiniMax open‑sources M2.7
MiniMax released M2.7 as an open model on Hugging Face, reporting state‑of‑the‑art scores on SWE‑Pro (56.22%) and Terminal Bench 2 (57.0%). (x.com) NVIDIA noted the model’s compatibility with GPU‑accelerated endpoints such as NemoClaw, and the release includes API support for developers. (x.com)
MiniMax has released M2.7 as an open model on Hugging Face, putting a 229 billion-parameter system aimed at coding and agent tasks into public download. (huggingface.co) The model card says M2.7 scored 56.22% on SWE-Pro and 57.0% on Terminal Bench 2, two tests built around software engineering and command-line work rather than short chatbot prompts. (huggingface.co) MiniMax describes M2.7 as a mixture-of-experts model, a design that keeps many specialist subnetworks on hand but activates only part of the system for each token; the company says the model has 256 experts, about 230 billion total parameters, and about 10 billion active per token. (developer.nvidia.com) That architecture is built for long, tool-using sessions instead of one-shot answers. MiniMax says M2.7 supports “interleaved thinking,” which means the model can pause between tool calls, inspect results, and choose its next step. (platform.minimax.io) MiniMax is pitching that behavior for “agent” software, which is shorthand for systems that can call tools, search files, and carry out multi-step jobs with limited supervision. Its product page says M2.7 is meant for “complex Agents” and long workflows with a 200,000-token context window. (minimax.io) The release is not just a model file drop. MiniMax’s API documentation lists M2.7 in its platform, and a separate guide positions it for coding tools with pay-as-you-go and token-plan access for developers who do not want to host the weights themselves. (platform.minimax.io, platform.minimax.io) NVIDIA has also moved quickly to make the model easier to run on its stack. A company blog post published April 11 says M2.7 works with NemoClaw, NVIDIA’s open-source setup for long-running assistants, and NVIDIA’s model catalog lists a hosted MiniMax M2.7 endpoint. (developer.nvidia.com, build.nvidia.com) The “open” label comes with caveats. Decrypt reported on April 14 that MiniMax changed commercial terms after release, while NVIDIA’s hosted listing says use of its endpoint is governed by NVIDIA trial terms and an open model license agreement. (decrypt.co, build.nvidia.com) What happens next is less about a benchmark screenshot than whether developers adopt M2.7 as a working engine for coding agents. MiniMax has now put the model on Hugging Face, in its own API, and on NVIDIA infrastructure, which gives it three distribution paths at once. (huggingface.co, platform.minimax.io, developer.nvidia.com)
