Hemmingway-1 MLX 4-bit

Uniform 4-bit quantization of Altworld/Hemmingway-1, a 27B parameter hybrid-attention language model built on Qwen3.8-27B.

Upstream model: Altworld/Hemmingway-1 by Altworld, released under Apache-2.0.

Conversion: Quantized to MLX format using Hermes Agent with mlx-lm (uniform 4-bit, group size 64).

Architecture

Component	Value
Parameters	27B
Hidden size	5120
Layers	64
Attention heads	24 (full) / 16 (linear)
KV heads	4 (full) / 48 (linear)
Intermediate size	17408
Vocab size	248,320
Context length	262,144
Attention type	Hybrid (linear + full every 4th layer)
RoPE theta	10,000,000

Quantization Details

Property	Value
Method	Uniform 4-bit
Bits per weight	4.501
Group size	64
Dtype	bfloat16

Quickstart

pip install -U mlx-lm

python3 -m mlx_lm.generate \
  --model hermitdave/Hemmingway-1-MLX-4bit \
  --prompt "Write a text to your landlord about the broken boiler." \
  --max-tokens 512 --temp 1.0 --top-p 0.95

Benchmark Results (DeepEval, local run)

Benchmarked on the oQ4e quant of this model (not this 4-bit quant) in a head-to-head vs Ornith-1.5-35B oQ4e, DeepEval GEval (LLM judge), run 2026-09-23, oMLX on M3 Max. 50 cases per model; message-not-memo sampled 3 generations per prompt.

Category	Hemmingway-1 oQ4e	Ornith-1.5 oQ4e
Directness	1.00	0.78
Human-Likeness	1.00	0.72
Empathy (EQ)	0.72	0.85
Hard asks — Actionability	0.64	0.53
Hard asks — Confidence	0.62	0.75
Message Not Memo	0.71	0.82
Story Quality	0.94	0.96

~5% of generations hit transient server errors and were excluded.

Citation

@misc{hemmingway1,
  title  = {Hemmingway-1},
  author = {Altworld},
  year   = {2026},
  url    = {https://huggingface.co/Altworld/Hemmingway-1},
}

License

Apache-2.0 (same as upstream).

hermitdave/Hemmingway-1-MLX-4bit