MiniMax and Unsloth ship 2-bit dynamic GGUF build of M3.1 Flash at 143.9GB
MiniMax announced on Oct 27 that M3.1 Flash has joined the Unsloth developer community, where Unsloth's dynamic GGUF quantization compresses the 428-billion-parameter mixture-of-experts model to a 143.9GB 2-bit footprint that runs locally in roughly 150GB of memory. In Unsloth's evaluation across seven domains, the quantized model retains 91.5 percent top-1 token agreement with the original on a token-by-token comparison, with Unsloth stating the build is 45 percent smaller while keeping 91.5 percent top-1 accuracy. Deployment runs through either Unsloth Desktop or llama.cpp, with an OpenAI-compatible API so existing applications and agents connect without interface changes. MiniMax framed the release around open weights, local control and more ways to build. Unsloth has more than 78,000 GitHub stars and over 500 million model downloads on Hugging Face, giving M3.1 Flash a path into an established community that already runs, evaluates and builds with open models. The 91.5 percent figure measures agreement with the original rather than task accuracy, and 2-bit quantization generally costs more on precision-sensitive work such as code and math, with real-world speed left for users to measure.