AgenticQwen-8B-oQ4

This repository contains an unofficial oMLX oQ4 quantized version of alibaba-pai/AgenticQwen-8B.

Original model card: https://huggingface.co/alibaba-pai/AgenticQwen-8B

What was done

The upstream alibaba-pai/AgenticQwen-8B safetensors model was downloaded from Hugging Face and quantized locally using the oMLX oQ streaming quantizer.

Quantization command used:

from omlx.oq import quantize_oq_streaming

quantize_oq_streaming(
    model_path="/Users/vilius/.omlx/models/alibaba-pai/AgenticQwen-8B",
    output_path="/Users/vilius/.omlx/models/alibaba-pai/AgenticQwen-8B-oQ4",
    oq_level=4,
    group_size=64,
    dtype="bfloat16",
)

The resulting model uses oMLX quantization metadata in config.json:

• default quantization: 4-bit affine
• group size: 64
• selected sensitive tensors retained at 5-bit or 6-bit by oQ
• output format: MLX/oMLX-compatible safetensors

Files

Main generated artifact:

model.safetensors

Tokenizer and generation/template files were copied from the upstream model:

added_tokens.json
chat_template.jinja
config.json
generation_config.json
merges.txt
special_tokens_map.json
tokenizer.json
tokenizer_config.json
vocab.json

Size

Approximate local sizes after conversion:

Model	Size
Original BF16 safetensors	~15 GB
oMLX oQ4 quantized	~4.5 GB

model.safetensors SHA256:

e5043b2e118c36ee43fa98b95ee155dfb90b0f9776bd43487154e7afdc053e70

Usage with oMLX

Download or place this repository under your oMLX models directory, then restart oMLX.

Example local layout:

~/.omlx/models/<namespace>/AgenticQwen-8B-oQ4

After restarting oMLX, the model should appear through the OpenAI-compatible models endpoint:

curl http://127.0.0.1:8000/v1/models

Expected model id if the folder is named AgenticQwen-8B-oQ4:

AgenticQwen-8B-oQ4

Example chat request:

curl http://127.0.0.1:8000/v1/chat/completions \
  -H 'Content-Type: application/json' \
  -d '{
    "model": "AgenticQwen-8B-oQ4",
    "messages": [
      {"role": "user", "content": "Write a short Python function to reverse a string."}
    ]
  }'

License

The upstream model is Apache 2.0. This quantized derivative is provided under the same license.

See the original model and its documentation here:

• alibaba-pai/AgenticQwen-8B
• Apache License 2.0

Notes

This is a quantized redistribution for convenience. It is not an official release from Alibaba PAI. For model details, intended use, training notes, and limitations, refer to the upstream model card.