AnkitAI/Parable-Qwen3-4B-Claude-Fable-5-MLX-8bit

Parable-Qwen3-4B-Claude-Fable-5 — MLX 8-bit

Apple Silicon build of Parable-Qwen3-4B-Claude-Fable-5, a Qwen3-4B fine-tuned on execution-verified agent traces.

~4.3 GB, 8 bits per weight. Larger and closer to the source than the 4-bit build; take this one if you have the RAM. Runs on any M-series Mac with room to spare.

Use it

pip install mlx-lm
mlx_lm.generate --model AnkitAI/Parable-Qwen3-4B-Claude-Fable-5-MLX-8bit \
  --prompt "Write a Python function that retries an HTTP call with backoff."

Or in Python:

from mlx_lm import load, generate
model, tokenizer = load("AnkitAI/Parable-Qwen3-4B-Claude-Fable-5-MLX-8bit")
print(generate(model, tokenizer, prompt="...", max_tokens=512))

What it is

Same weights as the source model, quantised to 4-bit for MLX. The recipe behind it is v3.1: LoRA on agent traces plus a replay mix, completion-only loss, two seeds souped, then merged into the base at scale 0.6 to limit drift.

Measured on the 4B, base against tuned in one session on one harness:

	base	v3.1
HumanEval+	0.616	0.683
MBPP+	0.603	0.638

Those numbers are from the full-precision model. Quantisation to 4 bits costs some accuracy; they are the ceiling, not a promise for this build.

Other formats

• GGUF — llama.cpp, LM Studio, Ollama
• safetensors — transformers

Apache-2.0, same as the base.