Qwen3.6-35B-A3B-NVFP4-GGUF
NVFP4 GGUF quantization of Qwen/Qwen3.6-35B-A3B.
Multimodal model: vision encoder (903 MB) + text MoE LLM (18.36 GB).
About NVFP4
NVFP4 is a Blackwell-native FP4 format (EFM4: 1 sign, 1 mantissa, 2 exponent bits). Applied uniformly to all tensors.
- Total params: 35.95B (3B active, 256 experts, 8/token)
- Quantization: NVFP4 (~4.55 BPW)
- File size: 18.36 GiB (text) + 903 MB (vision)
- Vision encoder: 27-layer ViT, hidden 1152, 3->1152x2 patch embed (temporal)
- Context: 262,144 tokens natively
NVFP4 requires a Blackwell (RTX 50-series or B-series) GPU for hardware acceleration.
Files
| Filename | Type | Size | Description |
|---|---|---|---|
qwen3.6-35b-a3b-nvfp4.gguf | NVFP4 | 18.36 GiB | Text MoE LLM weights |
mmproj-qwen36-35b-src-BF16.gguf | MMProj | 903 MB | Vision encoder weights |
Usage
llama.cpp CLI (text only)
llama-cli -hf FreedomAISVR/Qwen3.6-35B-A3B-NVFP4-GGUF -cnv -p "You are a helpful assistant"
llama-server (multimodal)
llama-server -hf FreedomAISVR/Qwen3.6-35B-A3B-NVFP4-GGUF --mmproj mmproj-qwen36-35b-src-BF16.gguf --ctx-size 0 --jinja
llama-cpp-python
from llama_cpp import Llama
llm = Llama.from_pretrained(
repo_id="FreedomAISVR/Qwen3.6-35B-A3B-NVFP4-GGUF",
filename="qwen3.6-35b-a3b-nvfp4.gguf",
)
Quantization Pipeline
# 1. Convert HF model to intermediate GGUF
python convert_hf_to_gguf.py ./models/qwen3.6-35b/ --outfile qwen3.6-35b-a3b-f16.gguf --outtype bf16
# 2. Export vision encoder
python convert_hf_to_gguf.py ./models/qwen3.6-35b/ --mmproj --outtype bf16
# 3. Quantize to NVFP4
llama-quantize --allow-requantize qwen3.6-35b-a3b-f16.gguf qwen3.6-35b-a3b-nvfp4.gguf NVFP4
Hardware
| GPU | VRAM | Notes |
|---|---|---|
| NVIDIA RTX 5060 Ti | 16 GB | Quantization performed on this GPU |
License
Apache-2.0 (same as Qwen/Qwen3.6-35B-A3B)