d3-nano-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-nano, the 2B multimodal decision model of Decision 3.0 from vLLM Semantic Router, built on Qwen/Qwen3.5-2B. 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 9.630 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	1.881B	84.981%	1.862 GiB	8.50
BF16	0.332B	14.996%	0.618 GiB	16.00
FP32	0.001B	0.024%	0.002 GiB	32.00
Total	2.214B	100%	2.482 GiB	9.630

Storage Breakdown

Component	Logical weights	Storage
Language backbone	1.882B	2.000 GB / 1.863 GiB
Vision encoder	0.331B	0.663 GB / 0.617 GiB
Code readout	0.001B	0.002 GB / 0.002 GiB
Total	2.214B	2.665 GB / 2.482 GiB

License

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