MiniMax M2.7 open‑source release
MiniMax published M2.7 as an open‑source model and reported top coding benchmarks—SWE‑Pro 56.22% and Terminal Bench 2 at 57.0%—and the model was made available on Ollama for commercial use. Ollama confirmed availability, enabling developers to run the model locally or via its platform. (x.com)
MiniMax has released M2.7 with downloadable weights, putting a new coding-focused model into the open model market on April 11 and 12, 2026. (github.com) (developer.nvidia.com) MiniMax’s own model card says M2.7 scored 56.22% on SWE-Pro and 57.0% on Terminal Bench 2, two tests built to measure how well a model handles software engineering tasks beyond simple code completion. (minimax.io) (huggingface.co) The model is also listed on Ollama, where developers can call `minimax-m2.7:cloud` with a 200,000-token context window through Ollama’s hosted service. Ollama’s library page shows the model was added about three weeks ago and had logged more than 57,000 downloads when it was crawled. (ollama.com) Large language models generate text by predicting the next token, or chunk of text, from patterns in training data. M2.7 uses a mixture-of-experts design, which is like routing each prompt to a small subset of specialists instead of waking up the whole model every time. (developer.nvidia.com) NVIDIA said the M2.7 family has 230 billion total parameters but activates 10 billion per token, with 256 experts and eight experts selected at a time. That setup is meant to cut inference cost while keeping the capacity of a much larger system available when needed. (developer.nvidia.com) MiniMax is pitching M2.7 as more than a chatbot for code snippets. Its GitHub and Hugging Face pages say the model was trained for “agent” work such as tool use, multi-step debugging, log analysis, root-cause checks, and long tasks spread across many files. (github.com) (huggingface.co) The company also says an internal version of M2.7 improved a programming scaffold over more than 100 rounds and produced a 30% performance gain. That claim comes from MiniMax’s own materials, and the company has not published an independent third-party audit alongside the release. (github.com) (huggingface.co) Access is open, but the license is not fully open-source in the standard commercial sense. The published license allows non-commercial use and says any commercial use requires prior written authorization from MiniMax. (huggingface.co 1) (huggingface.co 2) That leaves M2.7 in the growing “open weights” category: developers can download and run the model, but companies still need permission before turning it into a paid product or service. For now, the immediate shift is practical: M2.7 can be tested on common stacks including Ollama, Hugging Face, GitHub, and NVIDIA’s inference ecosystem. (developer.nvidia.com) (ollama.com) (github.com)
