MiniMax M3.1 Flash is now an open model that reads text, images and video

MiniMax has published the files of MiniMax M3.1 Flash, its model for coding, agentic workflows, computer use and cybersecurity, on Hugging Face. The model went live on October 19 through the MiniMax API, the M Plan subscription and the MiniMax Code desktop app, and the company's announcement that day said the model files would follow later. They went up on October 23. M3.1 Flash is now an open model that organizations can download and run on their own servers. Its license, however, sets tighter conditions than permissive licenses such as Apache 2.0; the details are below.

A model that reads text, images and video

According to the model card, M3.1 Flash is a natively multimodal mixture-of-experts (MoE) model: it takes text, image and video input and answers in text. It has about 428 billion parameters in total, 23 billion of which are active for each token, and a context window of up to 1 million tokens. MiniMax says an early version of the model was the anonymous model known as "Space Bunny".

MiniMax says M3.1 Flash keeps the core architecture of the earlier M3 with three changes. First, sparse attention now runs through the whole model: the first three layers, which used full attention in M3, have been switched to MiniMax Sparse Attention (MSA). Second, quantization-aware training: the expert layers and attention paths were trained at reduced precision, which the company says cuts memory use and makes serving more efficient. Third, DSpark speculative decoding: a small draft head guesses the next tokens and the main model checks them. The draft head ships with the model files. MiniMax says online inference is now nearly three times faster than before.

Model files and setup

The files on Hugging Face total about 250 gigabytes. A model of this size is beyond a personal computer; running it calls for data-center servers with several GPUs. MiniMax recommends SGLang, vLLM and TokenSpeed for serving, and says support for Transformers, KTransformers, unsloth and ATOM is in progress. A FlagOS guide covers running the model on different chip types. Reasoning depth can be set in five levels, from low to max. Usable context length and throughput depend on GPU memory and the serving setup.

License terms for commercial use

The model is released under MiniMax's own MiniMax Community License. Under the license text, non-commercial use is free. Commercial use comes with two conditions. Products or services that use the model must prominently display "Built with MiniMax M3.1-Flash" on a related website, interface or documentation. And if those products and services bring in more than 20 million US dollars a year, the user needs prior written authorization from MiniMax; below that threshold, a one-time notice to MiniMax is enough. The license counts paid services, commercial use of the model's API and commercial deployment of fine-tuned versions as commercial use. Its appendix of prohibited uses includes supporting any military purpose.

MiniMax's measurements

On the model card, MiniMax compares M3.1 Flash with DeepSeek V4.1 Flash, Kimi K3, GLM-5.3, GPT-5.6 Sol and Opus 5. On the software engineering benchmark DeepSWE v1.1, M3.1 Flash scores 73.7, while DeepSeek V4.1 Flash-1017 leads the table with 77.0 and Opus 5 scores 74.0. On OSWorld 2.0, which measures computer use, its 61.7 trails Opus 5 (70.6) and GPT-5.6 Sol (65.7). On the cybersecurity benchmark CyberGym it scores 82.2, behind GLM-5.3 (84.5) and GPT-5.6 Sol (83.6). It posts the highest score in the table on Automation-Bench with 62.8 and ties Opus 5 on BankerToolBench at 76.2.

MiniMax notes that the results apply to its official model files and that compressed versions made by others may score differently. All scores in the table are MiniMax's own measurements. For a fuller picture, follow reviews from independent researchers too.

Sources: Hugging Face (model card), MiniMax announcement, MiniMax release notes, License text