MiniMax M2.7 Open‑Sourced
MiniMax published an open‑source M2.7 model on Hugging Face and reported state‑of‑the‑art benchmark scores, including about 56.22% on SWE‑Pro and 57.0% on Terminal Bench 2. (x.com) The announcement included links to the model, a blog post and an API for developers to test. (x.com)
MiniMax has released M2.7 on Hugging Face, putting the weights for its latest coding-and-agent model into public download. (huggingface.co) The model card lists M2.7 as a 229 billion-parameter system and says it is aimed at “agent harnesses,” or software setups where a model can use tools, memory, and multi-step workflows instead of answering in one shot. (huggingface.co) MiniMax said M2.7 scored 56.22% on SWE-Pro, 57.0% on Terminal Bench 2, 55.6% on VIBE-Pro, and 39.8% on NL2Repo. The company also said the model reached 1495 Elo on GDPval-AA and 46.3% on Toolathon. (huggingface.co) Those tests are built to measure work closer to production software than short coding puzzles. SWE-Bench Pro describes itself as a benchmark for realistic enterprise software problems, and Terminal-Bench 2.0 says it evaluates agents on 89 sandboxed terminal tasks drawn from real workflows. (scaleapi.github.io) (arxiv.org) GDPval-AA measures models on knowledge-work tasks across 44 occupations and 9 industries, with shell access and web browsing in an agent loop. That puts MiniMax’s office-work claims in a benchmark category that goes beyond code generation. (artificialanalysis.ai) MiniMax published the M2.7 release on March 18, 2026, and its API release notes list two variants: MiniMax-M2.7 and M2.7-highspeed. The company’s product page says developers can try the model through its API and agent tools. (platform.minimax.io) (minimax.io) The company describes M2.7 as a model that took part in its own development cycle. In its blog post and model card, MiniMax said an internal version updated memory, built skills for reinforcement-learning experiments, and improved parts of its own harness based on results. (minimax.io) (huggingface.co) MiniMax said one internal M2.7 system optimized a programming scaffold for more than 100 rounds and produced a 30% performance improvement. That claim comes from MiniMax’s own materials, and the company did not publish an independent audit alongside the release. (huggingface.co) The open release also arrives as labs compete to show that open-weight models can do more than chat and code-complete. MiniMax’s earlier M2 release in October 2025 was positioned as a lower-cost “agentic” model with 230 billion total parameters and 10 billion active parameters. (huggingface.co) (platform.minimax.io) One detail developers will watch is licensing. The Hugging Face page for M2.7 lists a “modified-mit” license rather than plain MIT, while MiniMax links out to separate license terms from the repository page. (huggingface.co) For now, the release gives outside developers the model weights, an API endpoint, and a concrete set of benchmark claims to test. The next question is whether independent evaluations reproduce the numbers MiniMax posted on March 18. (huggingface.co) (platform.minimax.io)
