Foresee/Qwen3.8-9B-heretic-uncensored-5bit-MLX

🤗 Hugging Face sourcetext-generationapache-2.09B params18 GBsafetensorsHF checksums availableupdated today
No torrent yet

Qwen3.8-9B Heretic Uncensored - 5-bit MLX

This repository contains 5-bit MLX weights for rohit267/Qwen3.8-9B-heretic-uncensored.

The requested reference model was saga404/Qwen3.8-9B-heretic-uncensored-Q5_0-GGUF. MLX LM cannot convert GGUF weights directly. This conversion used the original BF16 safetensors weights to avoid an extra dequantization and requantization step.

Quantization

Setting Value
Weight bits 5
Group size 32
Mode affine
MLX LM version 0.31.3
MLX version 0.32.1

The final model uses about 6.001 bits per weight after scales, biases, and unquantized parameters are included.

Use

Install MLX LM:

pip install mlx-lm

Run the downloaded model:

mlx_lm.generate \
  --model ./Qwen3.8-9B-heretic-uncensored-5bit-MLX \
  --prompt "What is 2 + 2?" \
  --max-tokens 128

Use the source model's recommended sampling settings for longer responses:

temperature=0.6
top_p=0.95
top_k=20

Validation

The converted weights loaded and generated a correct response to 2 + 2 on Apple silicon. Peak memory during this smoke test was 6.859 GB.

License

The source model is licensed under Apache 2.0.