d3-mini-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-mini, the 4B multimodal decision model of Decision 3.0 from vLLM Semantic Router, built on Qwen/Qwen3.5-4B. 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 9.056 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	4.205B	92.618%	4.161 GiB	8.50
BF16	0.334B	7.367%	0.623 GiB	16.00
FP32	0.001B	0.014%	0.002 GiB	32.00
Total	4.540B	100%	4.786 GiB	9.056

Storage Breakdown

Component	Logical weights	Storage
Language backbone	4.206B	4.470 GB / 4.163 GiB
Vision encoder	0.334B	0.667 GB / 0.621 GiB
Code readout	0.001B	0.003 GB / 0.002 GiB
Total	4.540B	5.139 GB / 4.786 GiB

License

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