Qwen3.8-Flash-Next W4A16 NVFP4 GGUF
GGUF conversion of
axiomofmind/Qwen3.8-Flash-Next-W4A16-NVFP4,
based on Qwen/Qwen3.8-Flash-Next.
The routed-expert weights use W4A16 NVFP4. Attention, shared experts, routers, embeddings, PLE, and other retained tensors remain in BF16 or F32.
Files
| File | Description | Size |
|---|---|---|
Qwen3.8-Flash-Next-W4A16-NVFP4-BF16attn-PLE-noMTP.gguf | Main text model | 180.4 GB |
mmproj-Qwen3.8-Flash-Next-BF16.gguf | BF16 vision projector | 907.5 MB |
The vision projector is required for image inputs. It is not required for text-only use.
This GGUF does not include MTP weights.
Requirements
A llama.cpp build with Qwen4Exp and NVFP4 GGUF support is required.
For multimodal use, load the main model together with the included mmproj
file.
License
This model is distributed under the Qwen Community License 1.0. Refer to the official model card for architecture details, usage guidance, and limitations.