Qwen3.5-4B-MLX-8bit

Qwen3.5-4B-MLX-8bit#

This is a quantized MLX version of Qwen/Qwen3.5-4B for Apple Silicon.

Model Details#

• Original Model: Qwen/Qwen3.5-4B
• Quantization: 8-bit (~9.053 bits per weight)
• Group Size: 64
• Format: MLX SafeTensors
• Framework: mlx-vlm

Conversion Details#

This model was converted using mlx-vlm from the pc/fix-qwen35-predicate branch, which includes fixes for Qwen3.5 quantization predicates (proper handling of MoE gate layers, shared_expert_gate, and A_log casting).

Conversion command:

python3 -m mlx_vlm convert \
  --hf-path "Qwen/Qwen3.5-4B" \
  --mlx-path "./mlx_models/Qwen3.5-4B-MLX-8bit" \
  -q --q-bits 8 --q-group-size 64

Important Note#

A better, more optimized conversion may be available from @Prince (@Blaizzy) in the MLX VLM community. Check the mlx-community organization for updated versions as official Qwen3.5 support is merged into the main mlx-vlm branch.

Usage#

from mlx_vlm import load, generate

model, processor = load("mlx-community/Qwen3.5-4B-MLX-8bit")

output = generate(
    model,
    processor,
    prompt="Describe this image in detail",
    image="path/to/image.jpg",
    max_tokens=200
)
print(output)

Or from the command line:

mlx_vlm generate \
  --model mlx-community/Qwen3.5-4B-MLX-8bit \
  --prompt "Describe this image" \
  --image path/to/image.jpg \
  --max-tokens 200

Performance#

• Disk Size: ~4.8 GB
• Prompt Processing: ~173 tokens/sec (tested on Apple Silicon)
• Generation: ~88 tokens/sec (tested on Apple Silicon)
• Peak Memory: ~5.8 GB
• Runs efficiently on Apple Silicon Macs (M1/M2/M3/M4)

License#

This model inherits the Apache 2.0 license from the original Qwen3.5-4B model.