ultratensor models needs about 31.7 GB of memory at its main file (a 30.2 GB file) with an 8K context window, so it fits on a 32 GB GPU. This repository has 7 quants, from 17.1 GB to 102.4 GB. Longer context needs more: at 32K tokens the KV cache alone adds about 4.1 GB.
UltraTensor models: expert splices of DeepSeek-V4-Pro
UltraTensor models are spliced versions of DeepSeek-V4-Pro, keeping a subset of experts per MoE layer. The models are released in GGUF format and load with stock llama.cpp. They maintain the DeepSeek-V4 architecture and support a one million token context.
Good at
- Large context handling
- MoE expert specialization
- Stock llama.cpp compatibility
Limits
- No vision support
- Requires DeepSeek-V4 support in llama.cpp
- Large memory for full model
Best for
- Long‑form text generation
- Domain‑specific expert tasks
- Research on MoE pruning
Who it’s for. Developers and researchers who need a DeepSeek-V4-Pro variant with reduced expert count for faster inference or lower memory usage. The models are suitable for those who can run llama.cpp with DeepSeek-V4 support.
Quants & memory
| Quant | File | Est. memory | vs VRAM | Result | |
|---|---|---|---|---|---|
| — | 17.1 GB | 18.7 GB | Too big | Get file → | |
| — | 21.8 GB | 23.3 GB | Too big | Get file → | |
| — | 26.9 GB | 28.4 GB | Too big | Get file → | |
| — | 30.2 GB | 31.7 GB | Too big | Get file → | |
| — | 34.5 GB | 36.1 GB | Too big | Get file → | |
| — | 42.2 GB | 43.7 GB | Too big | Get file → | |
| — | 102.4 GB | 103.9 GB | Too big | Get file → |
The provided GGUF files are unquantized. For a 8 GB VRAM budget, the 17.14 GB file will need about 18.7 GB of VRAM at 8 k context, which is above the budget. A 21.82 GB file requires about 23.3 GB. For 16 GB VRAM, the 26.85 GB file needs about 28.4 GB. Users should consider quantizing the files to reduce memory usage.
What GPU do you need for ultratensor models?
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 the main file (21.8 GB file, about 23.3 GB in use).
- 32 GB VRAM: the best quant that fits is the main file (30.2 GB file, about 31.7 GB in use).
- 48 GB VRAM: the best quant that fits is the main file (42.2 GB file, about 43.7 GB in use).
Which ultratensor models quant should you choose?
If memory is tight, the smallest file is the first one in the table at 17.1 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 ultratensor models
How much VRAM does ultratensor models need?
About 31.7 GB at its main file with an 8K context window. That is the 30.2 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 ultratensor models on an 8 GB GPU?
Not at an 8K context. The smallest file needs about 18.7 GB, so you would need a larger GPU or CPU offloading.
What is the file size of ultratensor models GGUF?
It depends on the quant: file is 17.1 GB, file is 21.8 GB, file is 26.9 GB, file is 30.2 GB, file is 34.5 GB, file is 42.2 GB and 1 more.
What context length does ultratensor models support?
The GGUF header lists a trained context length of 1,048,576 tokens (1024K). The KV cache grows with the context you use: about 1.0 GB at 8K and 4.1 GB at 32K.
Where can I download ultratensor models GGUF?
The files are on Hugging Face in the NagusameCS/ultratensor-models repository. The download buttons on this page open the file on Hugging Face. ggufmodels.com does not host any model files.
What license does ultratensor models use?
Hugging Face lists the license as mit. 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 UltraTensor models on 8 GB VRAM?
The smallest file requires 18.7 GB of VRAM at 8 k context, so 8 GB is insufficient without quantization.
Do these models support vision tasks?
No, the models have no vision capabilities.
Which llama.cpp version is needed?
Stock llama.cpp with DeepSeek-V4 support is required; no fork is needed.