The user sixstringzen has released an enhanced oQ6e quantization of the Altworld/Hemmingway-1 model, optimized for MLX and oMLX on Apple silicon. This specific build is notable for preserving the model's multi-token prediction (MTP) tensors during the quantization process.
Model ReleasesHemmingway-1Altworld/Hemmingway-1
sixstringzen releases Hemmingway-1 oQ6e quantization with preserved MTP tensors for MLX
The conversion utilizes the oMLX 0.7.0.dev2 tool and employs an enhanced oQ6e mixed-precision affine quantization method. While the base precision is set to 6-bit with a group size of 64, the build assigns 8-bit precision to nine sensitive tensors, including the language_model.lm_head, to maintain higher accuracy. The resulting artifact has a size of approximately 21.39 GiB.
Structural and oMLX load checks confirmed the artifact's compatibility on 2026-09-20, with the model loading without format or weight errors. The weights are provided in MLX safetensors format. Users can load the model via the oMLX model browser, with the option to disable "enable_thinking" for direct prose generation.
Sources
- sixstringzen/Hemmingway-1-oQ6e-mtp（oMLX向けMTP保持6bit版） （Hugging Face: sixstringzen / Hemmingway-1 oQ6e, 2026-09-20）