Project Maya v1.1.30: 428B MiniMax M3.1 Flash quants fit a 128 GB PC
Local deployment: Maya-XS (93 GB) keeps every expert and Maya-S (122 GB) stays closest to the original, running on V100+ NVIDIA, Strix Halo and llama.cpp, plus a new loop guard against repeating answers.
Project Maya v1.1.30 is out: MiniMax-M3.1-Flash in our own quants (guffs), small enough for a 128 GB PC. Maya-XS (93 GB) holds the whole 428B model, every expert kept, in a 128 GB PC's memory. Maya-S (122 GB) is the one closest to the original. Both are made from MiniMax's own release and calibrated on what people actually use it for: code, tool calls and agents, front-end and three.js, reasoning, chat and other languages. What you get with them on Maya: - NVIDIA from V100 up and AMD, including Strix Halo, on Linux and Windows - Pictures and MiniMax's DSpark drafting, with the vision tower kept off the GPU until you send a picture - Prompt reading tuned to your PC once at setup, and the engine fitted to your memory from the first start - A new loop guard for every model on Maya: when an answer starts repeating itself, Maya steers it back out - The same files work in llama.cpp too Also in this release: a setup name for every quant (minimax-maya-xs, deepseek-maya-xs, glm-maya-s), DeepSeek with a 256K context by default where it fits, and fixes and improvements from our community. Thank you to everyone who sent a PR, tested or reported an issue. https://github.com/mw00/project-maya/releases/tag/v1.1.30