MiniMax open‑sources M2.7 model
MiniMax released its M2.7 model on Hugging Face and reports state‑of‑the‑art scores on developer benchmarks like SWE‑Pro (56.22%) and Terminal Bench 2 (57.0%), with NVIDIA highlighting the model for complex agentic workflows. The model is available with API access for experimentation and can be accelerated on NVIDIA hardware. (x.com)
MiniMax has released the weights for its M2.7 language model on Hugging Face, expanding public access to a system it says is built for coding and agent workflows. (huggingface.co) Large language models predict the next token, or next chunk of text, from patterns in training data; “mixture of experts” models split that work across specialized subnetworks so only part of the model runs on each step. NVIDIA said the MiniMax M2 series uses that sparse design to keep inference costs lower while preserving the capacity of a roughly 230 billion parameter model. (developer.nvidia.com) MiniMax describes M2.7 as a 229 billion parameter model aimed at software engineering, tool use, search, and office productivity, with features such as dynamic tool search and “Agent Teams” for multi-agent collaboration. The company’s model page says M2.7 is its “first model deeply participating in its own evolution.” (huggingface.co) The release is landing as open-weight models are being pushed beyond chatbots and into systems that can call tools, inspect codebases, and work through multi-step tasks. MiniMax’s own site says M2.7 is targeted at “complex Agents,” and NVIDIA framed the model around “complex agentic workflows” in software, machine learning research, and office work. (minimax.io, developer.nvidia.com) MiniMax is also using benchmark scores to place M2.7 in that race. The company says the model scored 56.22% on SWE-Pro, a software engineering benchmark, and 57.0% on Terminal Bench 2, which tests work done through a command-line environment rather than short code snippets. (marktechpost.com) Those claims are company-reported, and MiniMax’s Hugging Face card lists additional results including 46.3% on Toolathon and 62.7% on the MM Claw end-to-end benchmark. The same card says M2.7 reached an Elo score of 1495 on GDPval-AA, which it describes as the highest among open-source models. (huggingface.co) The model is not only downloadable. MiniMax’s developer docs list M2.7 through its API with token-plan and pay-as-you-go billing, and NVIDIA has published both a NIM model card and a deployable container for running the model on its hardware stack. (platform.minimax.io, build.nvidia.com, catalog.ngc.nvidia.com) That gives developers three routes to test it: download the open weights, call the hosted API, or deploy through NVIDIA’s inference tooling. For MiniMax, the bet is that M2.7 will be judged less by chatbot demos than by whether developers can make it finish real work. (huggingface.co, platform.minimax.io, developer.nvidia.com)
