In about ten days, independent developers reduced the open weights of MiniMax M3.1 Flash from 252.4 GB to roughly 93 GB, making the model runnable on far more modest machines. MiniMax itself thanked them by name in a post published the previous day.
M3.1 Flash is a mixture-of-experts model with about 428 billion total parameters, 23 billion active per token. It accepts text and image inputs, offers a context window of up to 1 million tokens, and is designed for coding, AI agents and computer-use tasks. The weights were released on October 23, after the online service launched on October 19, according to TechNode and the official model card.
Image: MiniMax — https://www.minimax.io/blog/minimax-m31f
Three compressed versions compared
The quantized versions cited by the community differ considerably in size. The smallest is Maya-XS by PeasantSmith, at 93 GB, which its author says is meant for a 128 GB machine. For tool-call and agent tests, the author reports 88.8% top-1 agreement.
A second version, SAGE-EXL3 by ViC305, takes up 108.64 GiB and targets a single DGX Spark. Its author reports 96.89% top-1 agreement on held-out sequences and decoding speeds between 24.37 and 29.72 tokens per second at a 48,000-token context.
The third, Unsloth UD-Q2_K_XL, weighs 143.9 GB and targets machines with more than 160 GB of memory. Its author reports 91.5% weighted top-1 agreement across seven categories.
Image: MiniMax — https://www.minimax.io/blog/minimax-m31f
The three figures come from different authors using different evaluation sets, so they are not directly comparable. They measure each compressed build against the original model, not an independent benchmark.
Why it matters
The story is about more than file size. A model that drops from over 250 GB to under 100 GB comes within reach of home workstations and small labs, which can run a large model locally rather than relying on an external service. For those handling sensitive data or high inference costs, that changes the calculation.
MiniMax's public thanks signals that the lab recognizes community work in distributing its open weights.
No comments yet. Be the first!