fable-traces-mxfp8-mlx

fable-traces — MLX Block float MX FP8

MLX quantization of AliesTaha/fable-traces, a fine-tuned Qwen3-4B-Instruct-2507 for short, conversational replies. This variant uses Block float MX FP8 quantization (8.25 effective bits/weight).

Quantized by: sahilchachra
Closest to FP16 quality; 8-bit block-float precision.

About the base model

• Architecture: Qwen3ForCausalLM — 36 layers, hidden 2560, 32 attention heads, 8 KV heads (GQA)
• Context length: 262 144 tokens
• Thinking mode: Qwen3 hybrid — supports <think> chain-of-thought with enable_thinking=True
• Fine-tune domain: Conversational / instruct (see egypt-won tag)
• License: Apache 2.0

Quick start

pip install mlx-lm

from mlx_lm import load, generate

model, tokenizer = load("sahilchachra/fable-traces-mxfp8-mlx")

messages = [{"role": "user", "content": "Tell me something interesting."}]
prompt = tokenizer.apply_chat_template(
    messages, tokenize=False, add_generation_prompt=True
)
response = generate(model, tokenizer, prompt=prompt, max_tokens=512, verbose=True)
print(response)

With thinking mode (Qwen3 chain-of-thought)

prompt = tokenizer.apply_chat_template(
    messages, tokenize=False, add_generation_prompt=True,
    enable_thinking=True,   # injects <think> block before answer
)
response = generate(model, tokenizer, prompt=prompt, max_tokens=1024, verbose=True)

CLI

mlx_lm.generate --model sahilchachra/fable-traces-mxfp8-mlx \
    --prompt "What's the fastest animal on Earth?" \
    --max-tokens 256

Quantization details

Note on bpw: Embedding and norm layers are kept at bf16; the reported bpw is across all linear weights.

All MLX variants

Credits

• Base fine-tune: AliesTaha/fable-traces by AliesTaha (Apache 2.0)
• Base architecture: Qwen/Qwen3-4B-Instruct-2507 by Qwen team
• MLX quantization by sahilchachra