peasantsmith/MiniMax-M3.1-Flash-Maya-GGUF

MiniMax-M3.1-Flash · Maya GGUF

GGUF quantizations of MiniMaxAI/MiniMax-M3.1-Flash (428B parameters, 23B active per token: 128 routed experts per layer, 4 active, plus a shared expert; sparse attention in every layer; up to a 1M-token context), made for Project Maya. Project Maya keeps the most-used experts on the GPU and computes the rest from RAM, so the model runs well below its size. Both quants are made from MiniMax's own release (FP4 experts, FP8 dense weights), not from a re-quantized file, and keep every expert.

Two quants

	Maya-S	Maya-XS
Made for	the closest answers to the original	the whole model in a 128 GB PC's memory
Model in memory	122.4 GB	93.0 GB
Routed experts	IQ2_XXS; the most sensitive layers' tensors IQ2_XS / IQ2_S / IQ3_XXS	IQ1_S; the most sensitive layers' tensors IQ2_XXS
Attention and sparse-attention indexer	Q8_0	Q5_K / Q6_K; indexer Q8_0
Shared experts, dense layers, embeddings, output	Q6_K	Q5_K / Q4_K / Q6_K
Next-token loss (NLL) vs the original	+0.8%	+5.0%

Quality: the same top token as the original

Category	Maya-S	Maya-XS
Tool calls and agents	92.1%	88.8%
three.js / WebGL / canvas	92.0%	88.5%
Web front-end	94.6%	91.7%
Code	93.7%	90.9%
Reasoning	85.9%	80.4%
Chat	76.4%	66.5%
Multilingual chat	78.5%	71.6%
Average of the categories	87.6%	82.6%

Measured on held-out conversations never used for calibration, on the model's own answers, against the original model's predictions (MiniMax's released weights, decoded exactly), with the files as published running in Project Maya's engine - the same results on NVIDIA and AMD.

How they were made

1. Calibration text in MiniMax's own chat format, thinking included: tool calls and agent traces, code (with three.js / canvas / WebGL and other front-end work), reasoning, long answers, chat, multilingual chat - 384 sequences of 2,048 tokens. A separate held-out set is kept for evaluation.
2. Statistics from the model itself, layer by layer: an importance matrix for every expert separately.
3. Where the bits go: every layer's gate, up and down matrices were measured in each candidate type, and the extra bits went to the layers where the model is most sensitive.
4. Error-feedback rounding of every expert, at the quantizer's 256-weight blocks, each expert against its own input statistics - and each layer calibrated on what the already-quantized layers before it produce.
5. llama.cpp's own quantizers (ggml), so the files are ordinary GGUFs in llama.cpp's minimax-m3 layout.

Files

Folder	Files	Size
Maya-S/	MiniMax-M3.1-Flash-Maya-S-IQ2_XXS-00001-of-00003.gguf … -00003-of-00003.gguf, and MiniMax-M3.1-Flash-Maya-S-maya-indexers.gguf	122.4 GB
Maya-XS/	MiniMax-M3.1-Flash-Maya-XS-IQ1_S-00001-of-00002.gguf, -00002-of-00002.gguf, and MiniMax-M3.1-Flash-Maya-XS-maya-indexers.gguf	93.0 GB
dspark/	MiniMax's DSpark drafter, as MiniMax released it	2.3 GB
vision/	MiniMax's vision tower, as MiniMax released it	1.7 GB

MiniMax-M3.1 runs its sparse-attention indexer in every layer. The indexers of layers 0-2 are in the small *-maya-indexers.gguf beside each quant: Project Maya loads them and runs the model as MiniMax defines it; stock llama.cpp loads the shards alone (it runs those three layers with full attention).

Running it

Project Maya sets everything up and runs it (v1.1.30 or later):

git clone https://github.com/mw00/project-maya.git && cd project-maya
./setup.sh --setup --model minimax-maya-s      # Windows: START-MAYA.bat
./setup.sh --setup --model minimax-maya-xs

It checks the PC (NVIDIA from V100 up, or AMD), downloads the files (every one verified by its sha256, resumed if a download stops), compiles the engine for your GPU(s) and starts it: the dashboard at http://127.0.0.1:8080, or the OpenAI- and Anthropic-compatible API at the same address (streaming, tool calls, pictures, MiniMax's thinking levels).

The shards also load in llama.cpp (minimax-m3).

License

The model is MiniMax's MiniMax-M3.1-Flash under the MiniMax Community License (LICENSE): free for non-commercial use; commercial use has conditions, including showing "Built with MiniMax M3.1-Flash". These quantizations are released under the same license.