DeepSeek V4 Flash Vision Strix Halo needs about 11.4 GB of memory at its main file (a 10.9 GB file) with an 8K context window, so it fits on a 12 GB GPU. This repository has 3 quants, from 10.9 GB to 91.5 GB.
DeepSeek V4 Flash Vision Strix Halo – multimodal model for AMD Strix Halo
DeepSeek V4 Flash Vision Strix Halo is a multimodal model built on the DeepSeek-V4-Flash architecture with a vision tower. It is designed to run only on AMD Strix Halo (gfx1151) using the Ember inference engine. The model includes a DSpark drafter and importance‑matrix‑calibrated quantization.
Good at
- Multimodal text and image understanding
- PNG image input only
- Ember inference on AMD Strix Halo
Limits
- Requires AMD Strix Halo (gfx1151) only
- Vision tower is F16 and not quantized
- Image input limited to PNG
Best for
- Vision‑enabled chatbot on AMD GPUs
- Image‑based data analysis on Ember
- Multimodal agent tasks with PNG images
Who it’s for. Developers and researchers needing a multimodal LLM that runs on AMD Strix Halo hardware. The model is suitable for applications that require both text and PNG image processing within the Ember inference framework.
Quants & memory
| Quant | File | Est. memory | vs VRAM | Result | |
|---|---|---|---|---|---|
| — | 10.9 GB | 11.4 GB | Too big | Get file → | |
| — | 10.9 GB | 11.4 GB | Too big | Get file → | |
| — | 91.5 GB | 92.1 GB | Too big | Get file → |
For a 10.9 GB file, the estimated VRAM requirement is 11.4 GB at 8 k context. This fits on a single 12 GB GPU. The 91.55 GB file requires about 92.1 GB VRAM, suitable for multi‑GPU setups. Choose the 10.9 GB quant if limited to a single GPU.
What GPU do you need for DeepSeek V4 Flash Vision Strix Halo?
These figures add the file size, an fp16 KV cache at an 8K context window (worked out from this model’s own header) and about 0.5 GB for the runtime. They are estimates, so leave a little headroom, and expect to use less context or a smaller quant if you are right at the limit.
- 4 GB VRAM: no quant fits at an 8K context.
- 6 GB VRAM: no quant fits at an 8K context.
- 8 GB VRAM: no quant fits at an 8K context.
- 12 GB VRAM: the best quant that fits is the main file (10.9 GB file, about 11.4 GB in use).
- 16 GB VRAM: the best quant that fits is the main file (10.9 GB file, about 11.4 GB in use).
- 24 GB VRAM: the best quant that fits is the main file (10.9 GB file, about 11.4 GB in use).
- 32 GB VRAM: the best quant that fits is the main file (10.9 GB file, about 11.4 GB in use).
- 48 GB VRAM: the best quant that fits is the main file (10.9 GB file, about 11.4 GB in use).
Which DeepSeek V4 Flash Vision Strix Halo quant should you choose?
If memory is tight, the smallest file is the first one in the table at 10.9 GB, with the quality trade-offs that come with fewer bits.
New to quant names? Read GGUF quantization explained.
Specifications
Architecture from the GGUF header
Source
Frequently asked questions about DeepSeek V4 Flash Vision Strix Halo
How much VRAM does DeepSeek V4 Flash Vision Strix Halo need?
About 11.4 GB at its main file with an 8K context window. That is the 10.9 GB file, plus the KV cache and roughly 0.5 GB of runtime overhead. Smaller quants need less and longer context needs more.
Can I run DeepSeek V4 Flash Vision Strix Halo on an 8 GB GPU?
Not at an 8K context. The smallest file needs about 11.4 GB, so you would need a larger GPU or CPU offloading.
What is the file size of DeepSeek V4 Flash Vision Strix Halo GGUF?
It depends on the quant: file is 10.9 GB, file is 10.9 GB, file is 91.5 GB.
Where can I download DeepSeek V4 Flash Vision Strix Halo GGUF?
The files are on Hugging Face in the otheru/DeepSeek-V4-Flash-Vision-Strix-Halo-GGUF repository. The download buttons on this page open the file on Hugging Face. ggufmodels.com does not host any model files.
What license does DeepSeek V4 Flash Vision Strix Halo use?
Hugging Face lists the license as other. Licenses can differ between the quantized files and the original model, so check the repository and the base model before commercial use.
Can I run this model on 8 GB VRAM?
No. The smallest quant requires 11.4 GB VRAM at 8 k context, so 8 GB is insufficient.
Does this model support JPEG images?
No. The vision tower accepts only non‑interlaced RGB/RGBA8 PNG images.
Is the model compatible with vLLM?
The model is designed for Ember on AMD Strix Halo; vLLM support is not provided.