d3 ultra needs about 18.5 GB of memory at Q4_K_M (a 15.8 GB file) with an 8K context window, so it fits on a 24 GB GPU. This repository has 4 quants, from 15.8 GB (Q4_K_M) to 27.3 GB (Q8_0). Longer context needs more: at 32K tokens the KV cache alone adds about 8.6 GB.
d3 ultra – 25.6B multimodal decision model
d3 ultra is a 25.6B parameter Qwen‑35 architecture model that reads text, JSON, images and videos once per request and answers multiple questions without generating text. It supports reasoning, tools and vision and offers a long context of 262144 tokens.
Good at
- Multimodal decision answering
- Single‑pass input processing
- High‑token context
Limits
- No text generation
- No vision projector file listed
- License is Apache‑2.0
Best for
- Batch question answering on images
- Video analysis with bounding boxes
- Rapid decision making in APIs
Who it’s for. Developers needing fast, multimodal decision logic without text generation. Suitable for applications that require single‑pass processing of images or videos and multiple question types.
Quants & memory
| Quant | File | Est. memory | vs VRAM | Result | |
|---|---|---|---|---|---|
| Q4_K_M | 15.8 GB | 18.5 GB | Too big | Get file → | |
| Q5_K_M | 18.4 GB | 21.0 GB | Too big | Get file → | |
| Q6_K | 21.0 GB | 23.7 GB | Too big | Get file → | |
| Q8_0 | 27.3 GB | 29.9 GB | Too big | Get file → |
For 8 GB VRAM, use Q4_K_M (15.84 GB file, 18.5 GB memory at 8 k context). For 12 GB VRAM, Q5_K_M (18.36 GB file, 21 GB memory). For 16 GB VRAM, Q6_K (21.04 GB file, 23.7 GB memory).
What GPU do you need for d3 ultra?
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: no quant fits at an 8K context.
- 16 GB VRAM: no quant fits at an 8K context.
- 24 GB VRAM: the best quant that fits is Q6_K (21.0 GB file, about 23.7 GB in use).
- 32 GB VRAM: the best quant that fits is Q8_0 (27.3 GB file, about 29.9 GB in use).
Which d3 ultra quant should you choose?
Q4_K_M is the usual starting point for d3 ultra: it balances file size and quality well. Here it is a 15.8 GB file.
The Q8_0 file is 27.3 GB, about 1.7 times the size of Q4_K_M. It keeps the weights closer to the original, which only matters if you have the memory to spare.
New to quant names? Read GGUF quantization explained.
Specifications
Architecture from the GGUF header
Source
Frequently asked questions about d3 ultra
How much VRAM does d3 ultra need?
About 18.5 GB at Q4_K_M with an 8K context window. That is the 15.8 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 d3 ultra on an 8 GB GPU?
Not at an 8K context. The smallest file needs about 18.5 GB, so you would need a larger GPU or CPU offloading.
What is the file size of d3 ultra GGUF?
It depends on the quant: Q4_K_M is 15.8 GB, Q5_K_M is 18.4 GB, Q6_K is 21.0 GB, Q8_0 is 27.3 GB.
What context length does d3 ultra support?
The GGUF header lists a trained context length of 262,144 tokens (256K). The KV cache grows with the context you use: about 2.1 GB at 8K and 8.6 GB at 32K.
Where can I download d3 ultra GGUF?
The files are on Hugging Face in the magnitudedev/d3-ultra-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 d3 ultra use?
Hugging Face lists the license as apache-2.0. 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 d3 ultra on 8 GB VRAM?
Yes, using the Q4_K_M quantized file which requires about 18.5 GB of VRAM at 8 k context.
Does d3 ultra generate text?
No, it answers questions with probabilities, choices, scores or bounding boxes without generating text.