clef-flash-oQ4e
Updated 2026-10-29: Re-quantized with the improved oQe quantization in oMLX 0.7.3. Please re-download if you have an earlier copy.
This model was quantized using oQ (oMLX v0.7.3) mixed-precision quantization with improved oQe (jundot/omlx#4930).
A mixed-precision MLX checkpoint of Cloudflare/clef-flash, a 9B decision model post-trained from Qwen/Qwen3.5-9B that turns a state and a schema of typed questions into decisions, returning a probability for every allowed option of every question in a single forward pass. It includes the Qwen3.5-9B language backbone, the vision encoder and the joint schema head.
The joint schema head (joint_head.safetensors, joint_head_config.json) is copied unchanged from the source. oMLX serves this checkpoint through POST /v1/systemone.
The effective average is 4.842 bits/weight for the language backbone and 5.518 bits/weight for the complete checkpoint.
Quantization and Bit Distribution
oQ4e allocates bits by measured layer sensitivity, then rounds each group with an importance matrix (128 samples x 512 tokens) and a weighted least-squares refit of its scale and bias.
| Storage format | Logical weights | Share of total | Tensor storage | Effective bits/weight |
|---|---|---|---|---|
| Affine 4-bit, group 64 | 6.924B | 72.640% | 3.627 GiB | 4.50 |
| Affine 5-bit, group 64 | 1.012B | 10.616% | 0.648 GiB | 5.50 |
| Affine 6-bit, group 64 | 1.017B | 10.671% | 0.770 GiB | 6.50 |
| BF16 | 0.579B | 6.073% | 1.078 GiB | 16.00 |
| Total | 9.532B | 100% | 6.123 GiB | 5.518 |
Storage Breakdown
| Component | Logical weights | Storage |
|---|---|---|
| Language backbone | 8.954B | 5.419 GB / 5.047 GiB |
| Vision encoder | 0.456B | 0.912 GB / 0.849 GiB |
| Joint schema head | 0.122B | 0.244 GB / 0.227 GiB |
| Total | 9.532B | 6.574 GB / 6.123 GiB |
License
Apache-2.0, following the source model. See LICENSE.
- Downloads last month
- 1,670
4-bit
# Make sure mlx-vlm is installed # pip install --upgrade mlx-vlm from mlx_vlm import load, generate from mlx_vlm.prompt_utils import apply_chat_template from mlx_vlm.utils import load_config # Load the model model, processor = load("Jundot/clef-flash-oQ4e") config = load_config("Jundot/clef-flash-oQ4e") # Prepare input image = ["http://images.cocodataset.org/val2017/000000039769.jpg"] prompt = "Describe this image." # Apply chat template formatted_prompt = apply_chat_template( processor, config, prompt, num_images=1 ) # Generate output output = generate(model, processor, formatted_prompt, image) print(output)