You can deploy this model and start earning money today!

dylpickle13/
Qwen3.5-0.8B-oQ8-mtp

Name: Qwen3.5-0.8B-oQ8-mtp
Availability: InStock

Views

No views yet

Qwen3.5-0.8B-oQ8-mtp

This is an oMLX/MLX-compatible 8-bit oQ quantization of Qwen/Qwen3.5-0.8B, with VLM support and native MTP tensors preserved.

This is not an official Qwen, Alibaba, MLX, or oMLX release.

Quantization details

Quantized with oMLX 0.3.9 using oQ quantization:

json

{
  "base_model": "Qwen/Qwen3.5-0.8B",
  "oq_level": 8,
  "group_size": 64,
  "text_only": false,
  "dtype": "bfloat16",
  "preserve_mtp": true
}

Intended use

Low-latency local vision-language inference
Camera/dashboard observation
Edge VLM experiments on Apple Silicon with oMLX/MLX

Compatibility notes

Verified locally with oMLX:

Model type: qwen3_5
Engine type: vlm
VLM capable: yes
Native MTP compatible: yes
MTP heads preserved: yes
OpenAI-compatible text chat endpoint: passed
OpenAI-compatible image chat endpoint: passed

Smoke tests

The local build was tested with:

Text prompt: Reply with exactly: OK
- Expected: OK
- Result: passed
Image prompt: a generated red square with What is the dominant color? Reply with one lowercase word.
- Expected: red
- Result: passed

License

The base model Qwen/Qwen3.5-0.8B is distributed under Apache 2.0 according to its Hugging Face model metadata. This quantized derivative is released under the same license.

Please review and comply with the base model license and any applicable acceptable-use guidance from the original authors.

dylpickle13/Qwen3.5-0.8B-oQ8-mtp