Maya-M is out: the larger quant (116 GB), closer to the full model token by token - and 97.9% of the FP8 model's zero-shot accuracy, like Maya-S. Update with git pull; set Maya-M up with ./setup.sh --setup --model Maya-M (Windows: START-MAYA.bat --setup --model Maya-M).
What's new
- Maya-M on Hugging Face (
Maya-M/): IQ2_S gate/up experts with error-feedback rounding, IQ3_XXS down
projections, IQ3_S in the most sensitive layers, from Z.ai's FP8 release with the FP8 model's own statistics,
calibrated toward tool calls and front-end code. Against the FP8 model: 23% lower KL divergence than Maya-S
(0.329 vs 0.428), the same next token 86.2% of the time (Maya-S 83.3%), and 97.9% of its zero-shot accuracy - the
same as Maya-S (multiple-choice tasks do not separate the two). Set it up with
./setup.sh --setup --model Maya-M(Windows:START-MAYA.bat --setup --model Maya-M); downloads are checked
against their published sha256. Maya-S stays the default. Results:bench/results/MAYA-M.md. - Decode: the dense matrix-vector products run in kernels compiled for their weight type (Q6_K, Q8_0, Q4_K,
Q5_K) - 10% faster per product, bit-identical results; it shows on cards that hold most experts in VRAM. --modelnow takes that download even when an earlier setup used--gguf-dir.