d3-flash-oQ8e

This model was quantized using oQ (oMLX v0.7.1) mixed-precision quantization with Improved oQe Quantization (jundot/omlx#4385).

A mixed-precision MLX checkpoint of vllm-sr/d3-flash, the 9B multimodal decision model of Decision 3.0 from vLLM Semantic Router, built on Qwen/Qwen3.5-9B. It reads text or JSON with optional images and videos and returns a probability for every option of choice, yes/no and score questions without generating text. It includes the language backbone, the vision encoder and the code readout.

The code readout (readout.safetensors, FP32) and decision_config.json are copied unchanged from the source. oMLX serves this checkpoint through POST /v1/systemone; Decision 3.0 support (jundot/omlx#4470) is newer than oMLX v0.7.1.dev1.

The effective average is 8.911 bits/weight for the complete checkpoint.

Quantization and Bit Distribution

oQ8e 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 8-bit, group 64	7.936B	94.542%	7.853 GiB	8.50
BF16	0.457B	5.445%	0.851 GiB	16.00
FP32	0.001B	0.012%	0.004 GiB	32.00
Total	8.394B	100%	8.708 GiB	8.911

Storage Breakdown

Component	Logical weights	Storage
Language backbone	7.937B	8.434 GB / 7.855 GiB
Vision encoder	0.456B	0.912 GB / 0.849 GiB
Code readout	0.001B	0.004 GB / 0.004 GiB
Total	8.394B	9.350 GB / 8.708 GiB

License

Apache-2.0, following the source model. See LICENSE and NOTICE.