Downloads · 30 days
148
29% of all-time downloads
Xananthium/Qwen3.6-35B-A3B-Hauhau-Aggressive-W4A16-Vision
Qwen3.6-35B-A3B-Hauhau-Aggressive-W4A16-Vision is a image-text-to-text model from Xananthium. Use it for the image-text-to-text task on the model card, and read the license before you ship it in a product. It is set up for transformers. The card lists the license as apache-2.0.
This is a vLLM-ready, multimodal W4A16 checkpoint built from the aggressive Hauhau Qwen3.6 language model while retaining the official Qwen3.6 vision tower and multimodal configuration.
Downloads · 30 days
148
29% of all-time downloads
All-time downloads
515
Public
Parameters
6.8B
22.2 GB on disk
Likes
0
Public
Click a slice to open those files.
.safetensors22.1 GB · 100%
How the weights are stored.
I324.2B · 62%
From the Hugging Face model README
This is a vLLM-ready, multimodal W4A16 checkpoint built from the aggressive Hauhau Qwen3.6 language model while retaining the official Qwen3.6 vision tower and multimodal configuration.
The checkpoint is approximately 21 GiB and uses symmetric 4-bit weights with
16-bit activations and group size 128. It is packed as
auto_round:auto_gptq; it is not AWQ.
HauhauCS/Qwen3.6-35B-A3B-Uncensored-HauhauCS-Aggressive,
using its highest-fidelity Q8_K_P GGUF artifact.Qwen/Qwen3.6-35B-A3B.auto_round:auto_gptq packing.The GGUF language tensors were dequantized, inverse Qwen3.6 layout transforms were applied, and the result was reconstructed into the native Hugging Face tensor layout. The official BF16 vision tower was then retained before the language checkpoint was packed to W4A16.
See reconstruction_metadata.json and
scripts/reconstruct_hauhau_qwen36.py for provenance and reconstruction
details.
The uploaded files are the same checkpoint used for the following local tests:
The tested deployment used two GPUs and a 180,000-token serving window:
vllm serve Xananthium/Qwen3.6-35B-A3B-Hauhau-Aggressive-W4A16-Vision \
--served-model-name qwen3.6-35b-hauhau-w4a16-vision \
--tensor-parallel-size 2 \
--dtype bfloat16 \
--gpu-memory-utilization 0.94 \
--max-model-len 180000 \
--max-num-seqs 3 \
--max-num-batched-tokens 32768 \
--enable-prefix-caching \
--enable-auto-tool-choice \
--tool-call-parser qwen3_coder \
--reasoning-parser qwen3 \
--trust-remote-code
The local deployment also used 4-bit TurboQuant KV cache and 16 GiB native CPU KV offload. Those are runtime choices, not requirements of the checkpoint:
--kv-cache-dtype turboquant_4bit_nc \
--kv-offloading-backend native \
--kv-offloading-size 16
For smaller GPUs, reduce --max-model-len, concurrency, or GPU memory
utilization. Runtime support for this packing format is required; the tested
stack was vLLM 0.27.1 with Transformers 5.15.1.
When served with the tested vLLM release, the model was validated through both OpenAI-compatible and Anthropic-compatible routes, including Anthropic image content blocks and native tool calls. This makes it suitable for Claude Code behind an authenticated reverse proxy.
This derivative is distributed under Apache-2.0, matching the source checkpoint. Credit belongs to the Qwen team for Qwen3.6 and HauhauCS for the aggressive language-model source. Review the source repositories for their full model cards and intended-use notes.
Not evaluated in the 2026-10-06 matched precision comparison. The attached report describes other checkpoints, not scores for this artifact.
See results/precision-comparison.json for measured results and limitations.