I converted and published six Hemmingway-1 builds for MLX and oMLX on Apple Silicon, with model cards that describe the source revision, quantization choices, calibration, and verification. Altworld developed the source model. My work concerns making those weights usable locally and examining the tradeoffs.
The collection spans oQ2e through oQ8e. Its companion preference study compares a fixed writing suite with a local BF16 reference. A separate evidence package examines BF16-to-oQ4e numerical fidelity and records controlled oQ4e runtime behavior. Each answers a different question, and the published limits keep those conclusions separate.
The public record includes a correction to the first aggregate analysis. I want someone reading the work to be able to see what changed, inspect the supporting material, and understand what still needs to be tested.
Explore the model cards, studies, and correction record through the evidence links below.
Supporting evidence
- Bobby Pierce on Hugging Face
- Hugging Face model collection: The upstream model, six converted builds, and companion evidence packages.
- Corrected preference study: Fixed writing suite, selected execution metadata, and corrected aggregate judgments.
- Fidelity and runtime evidence: BF16-to-oQ4e fidelity and single-condition oQ4e runtime measurements.
- Analysis correction: The September 22 correction and its effect on the aggregate analysis.