Converted and measured on one machine — Apple M3 Ultra, 96 GB unified memory, macOS 27 — as part of a full ladder. Every rung in this family came from the same BF16 source with the same group size, so bit width is the only variable between them.


Size on disk	27.99 GB
Perplexity	105.975
Relative to best rung in family	1.06×
Throughput (1 req / 8 concurrent)	43.8 / 143.8 tok/s

Perplexity measured on allenai/tulu-3-sft-mixture, 192 samples of 512 tokens, seed 123 — identical for every rung.

Perplexity is only comparable within this family. Tokenizers differ between model families, so a number here should never be compared against a different base model's. The × column above is the meaningful one.

Usage

bash

pip install mlx-vlm
mlx_vlm.generate --model shoemoney/Gemma-4-26B-A4B-Heretic-MLX-q8 --prompt "Hello" --max-tokens 256

Load with mlx-vlm, not mlx-lm — this architecture is registered in mlx-vlm.

Provenance

bash

mlx_vlm.convert --hf-path coder3101/gemma-4-26B-A4B-it-heretic \
                --mlx-path Gemma-4-26B-A4B-Heretic-q8 -q --q-bits 8 --q-group-size 64

License

apache-2.0, inherited from the base model. Attribution above.