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

Apple Silicon build of Parable-Qwen3-4B: 2.1 GB at 4.501 bits per weight, running natively on MLX with no llama.cpp in the way.

A 4-bit MLX quantisation of AnkitAI/Parable-Qwen3-4B-Claude-Fable-5, a Qwen3-4B fine-tune trained on real multi-step agent sessions: planning, tool use, and <think> reasoning captured from actual Claude Fable 5 and GPT-5.5 agent work, not synthetic Q&A. Fits comfortably on any M-series Mac.

Usage

pip install mlx-lm

mlx_lm.generate --model AnkitAI/Parable-Qwen3-4B-Claude-Fable-5-MLX-4bit \
  --prompt "Write a Python function that retries an HTTP request with exponential backoff."

Or from Python:

from mlx_lm import load, generate

model, tokenizer = load("AnkitAI/Parable-Qwen3-4B-Claude-Fable-5-MLX-4bit")
messages = [{"role": "user", "content": "Write a Python function that retries an HTTP request with exponential backoff."}]
prompt = tokenizer.apply_chat_template(messages, add_generation_prompt=True)
print(generate(model, tokenizer, prompt=prompt, max_tokens=512))

Recipe

v3.1: LoRA on agent traces with a replay mix to limit forgetting, completion-only loss so the model trains on answers rather than prompts, two seeds souped, then merged into the base at scale 0.6 to bound drift from the original weights.

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

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

Those are the full-precision numbers. Quantising to 4 bits costs accuracy that this table does not measure, so treat them as the ceiling for this build rather than a claim about it.

Other formats

format	repo	for
GGUF	Parable-Qwen3-4B-Claude-Fable-5-GGUF	llama.cpp, LM Studio, Ollama
MLX 8-bit	Parable-Qwen3-4B-Claude-Fable-5-MLX-8bit	Apple Silicon, closer to source
safetensors	Parable-Qwen3-4B-Claude-Fable-5	transformers

Apache-2.0, inherited from the base model.