MiniMax-M3

MiniMax-M3 is an upcoming AI language model from MiniMax, the company behind MiniMax-M2.7. At the time this repository was created, M3 had not been released yet. The repo exists as a community feedback hub, not as a working codebase. MiniMax is using it to collect user reports before finalizing the new model.

The team is specifically looking for feedback on the current model, M2.7. They want to hear about bugs or unexpected outputs, tasks that M2.7 still handles poorly, performance gaps compared to other benchmarks, and pain points developers have run into when deploying M2.7 using tools like SGLang, vLLM, or the Transformers library. They are also interested in feedback from people who have built automated workflows or custom tools on top of M2.7.

Feedback can be sent through several channels. GitHub Issues is the main route for bug reports and capability requests, with structured templates provided. A WeChat group serves Chinese-speaking users, a Discord server is available for English speakers, and a direct email address is listed for private or partnership inquiries. If you are reporting a bug, the README asks you to include which inference path you used, the settings you ran the model with, and a minimal example showing the problem.

While M3 is being developed, M2.7 remains the latest available model. It can be accessed through the MiniMax API, their web Agent product, or downloaded and run locally using standard open-source inference tools. The recommended generation settings for M2.7 are also noted in the README.

This repository contains no source code and no downloadable model weights for M3. It is a placeholder and feedback channel. Anyone interested in following the M3 release can watch the repository for future announcements.