Deepseek-V3.1-Terminus needs about 407.0 GB of memory at Q4_K_M (a 405.4 GB file) with an 8K context window, which is more than a 48 GB GPU holds. This repository has 27 quants, from 163.2 GB (TQ1_0) to 1342.3 GB (BF16). Longer context needs more: at 32K tokens the KV cache alone adds about 4.3 GB.
Deepseek‑V3.1‑Terminus – 670 billion‑parameter MLP‑O‑model
Deepseek‑V3.1‑Terminus is a 670 billion‑parameter deepseek2 architecture with a 163 840‑token context window. It is released under the MIT license and supports chat, code, and search agents. The model is available in multiple GGUF quantizations for llama.cpp.
Good at
- Large context handling
- Agent‑based tasks
- High‑parameter depth
Limits
- Requires 8 k context for memory estimates
- No vision support
- Large file sizes for high‑quality quants
Best for
- Long‑form conversation
- Complex code generation
- Search‑enabled dialogue
Who it’s for. Developers and researchers needing a high‑capacity language model with extensive context. Suitable for local deployment on systems with sufficient VRAM and storage.
Quants & memory
| Quant | File | Est. memory | vs VRAM | Result | |
|---|---|---|---|---|---|
| TQ1_0 | 163.2 GB | 164.8 GB | Too big | Get file → | |
| IQ1_S | 186.7 GB (4 parts) | 188.2 GB | Too big | Get file → | |
| IQ1_M | 201.0 GB (5 parts) | 202.6 GB | Too big | Get file → | |
| IQ2_XXS | 217.1 GB (5 parts) | 218.7 GB | Too big | Get file → | |
| IQ2_M | 228.9 GB (5 parts) | 230.5 GB | Too big | Get file → | |
| Q2_K | 245.7 GB (5 parts) | 247.3 GB | Too big | Get file → | |
| Q2_K_L | 245.9 GB (5 parts) | 247.5 GB | Too big | Get file → | |
| Q2_K_XL | 251.3 GB (6 parts) | 252.9 GB | Too big | Get file → | |
| IQ3_XXS | 273.5 GB (6 parts) | 275.1 GB | Too big | Get file → | |
| Q3_K_S | 290.4 GB (6 parts) | 292.0 GB | Too big | Get file → | |
| Q3_K_XL | 296.2 GB (7 parts) | 297.8 GB | Too big | Get file → | |
| Q3_K_M | 320.5 GB (7 parts) | 322.0 GB | Too big | Get file → | |
| IQ4_XS | 358.3 GB (8 parts) | 359.9 GB | Too big | Get file → | |
| IQ4_NL | 379.1 GB (8 parts) | 380.7 GB | Too big | Get file → | |
| Q4_0 | 380.1 GB (8 parts) | 381.7 GB | Too big | Get file → | |
| Q4_K_S | 381.1 GB (8 parts) | 382.7 GB | Too big | Get file → | |
| Q4_K_XL | 384.3 GB (8 parts) | 385.9 GB | Too big | Get file → | |
| Q4_K_M | 405.4 GB (9 parts) | 407.0 GB | Too big | Get file → | |
| Q4_1 | 420.8 GB (9 parts) | 422.4 GB | Too big | Get file → | |
| Q5_K_S | 462.6 GB (10 parts) | 464.2 GB | Too big | Get file → | |
| Q5_K_M | 476.2 GB (10 parts) | 477.7 GB | Too big | Get file → | |
| Q5_K_XL | 481.9 GB (10 parts) | 483.4 GB | Too big | Get file → | |
| Q6_K | 551.3 GB (12 parts) | 552.9 GB | Too big | Get file → | |
| Q6_K_XL | 574.9 GB (12 parts) | 576.5 GB | Too big | Get file → | |
| Q8_0 | 713.3 GB (15 parts) | 714.9 GB | Too big | Get file → | |
| Q8_K_XL | 780.6 GB (17 parts) | 782.2 GB | Too big | Get file → | |
| BF16 | 1342.3 GB (30 parts) | 1343.9 GB | Too big | Get file → |
For 8 GB VRAM, use Q4_0 (380.1 GB file, 381.7 GB memory). For 16 GB VRAM, Q4_K_S (381.06 GB file, 382.7 GB memory). For 32 GB VRAM, Q4_K_M (405.4 GB file, 407 GB memory).
What GPU do you need for Deepseek-V3.1-Terminus?
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: no quant fits at an 8K context.
- 32 GB VRAM: no quant fits at an 8K context.
- 48 GB VRAM: no quant fits at an 8K context.
Which Deepseek-V3.1-Terminus quant should you choose?
Q4_K_M is the usual starting point for Deepseek-V3.1-Terminus: it balances file size and quality well. Here it is a 405.4 GB file.
The Q8_0 file is 713.3 GB, about 1.8 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.
If memory is tight, the smallest file is TQ1_0 at 163.2 GB, with the quality trade-offs that come with fewer bits. Lower bit counts mean smaller files and more quality loss.
New to quant names? Read GGUF quantization explained.
Specifications
Architecture from the GGUF header
Source
Frequently asked questions about Deepseek-V3.1-Terminus
How much VRAM does Deepseek-V3.1-Terminus need?
About 407.0 GB at Q4_K_M with an 8K context window. That is the 405.4 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-V3.1-Terminus on an 8 GB GPU?
Not at an 8K context. The smallest file needs about 164.8 GB, so you would need a larger GPU or CPU offloading.
What is the file size of Deepseek-V3.1-Terminus GGUF?
It depends on the quant: TQ1_0 is 163.2 GB, IQ1_S is 186.7 GB, IQ1_M is 201.0 GB, IQ2_XXS is 217.1 GB, IQ2_M is 228.9 GB, Q2_K is 245.7 GB and 21 more.
What context length does Deepseek-V3.1-Terminus support?
The GGUF header lists a trained context length of 163,840 tokens (160K). The KV cache grows with the context you use: about 1.1 GB at 8K and 4.3 GB at 32K.
Where can I download Deepseek-V3.1-Terminus GGUF?
The files are on Hugging Face in the unsloth/DeepSeek-V3.1-Terminus-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-V3.1-Terminus 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 Deepseek‑V3.1‑Terminus on 8 GB VRAM?
Use the Q4_0 quant; it requires approximately 381.7 GB of VRAM at 8 k context, which is not feasible on 8 GB. Lower‑rank quants are needed.
Does the model support vision?
No, the model has no vision capabilities.
What license applies to the model?
The model is licensed under the MIT license.