You need to agree to share your contact information to access this model

This repository is publicly accessible, but you have to accept the conditions to access its files and content.

Qwen3.5-9B-NVFP4-QAD-LR1e-5-s4000-weight-only

Weight-only (W4A16 serving contract) export of the same trained checkpoint as Qwen3.5-9B-NVFP4-QAD-LR1e-5-s4000 — see that card for the full training provenance (recipe, dataset, W&B) and the measured 2x2 tables. The A16 rows in those tables were measured on this artifact.

Serving

compressed-tensors NVFP4 weight-only (W4A16) artifact: same trained weights as Qwen3.5-9B-NVFP4-QAD-LR1e-5-s4000, exported with scripts/export_nvfp4_vllm.py --weight-only — input_activations is null and no input_global_scale tensors are shipped, so engines serve it W4A16 with no activation quantization anywhere. This is the exact artifact behind the A16 rows in the tables.

vllm serve weili-0234/Qwen3.5-9B-NVFP4-QAD-LR1e-5-s4000-weight-only --max-model-len 24576

How this artifact was produced

Exported from the arm's final training checkpoint with scripts/export_nvfp4_vllm.py --weight-only (QATFactory weili/w4a4 @ a9315e4, PR). Training provenance is identical to Qwen3.5-9B-NVFP4-QAD-LR1e-5-s4000; the BF16 master weights live at Qwen3.5-9B-NVFP4-QAD-LR1e-5-s4000-BF16.

Part of a monitored QAD experiment series with full bookkeeping (pre-registered predictions, exact SHAs/configs/seeds per run). Produced with AI assistance (Claude).

Downloads last month: -

Safetensors

Model size

6B params

Tensor type

F32

BF16

F8_E4M3

Model tree for weili-0234/Qwen3.5-9B-NVFP4-QAD-LR1e-5-s4000-weight-only

Base model

Qwen/Qwen3.5-9B-Base

Finetuned

Qwen/Qwen3.5-9B

Quantized

(479)

this model