What Do You Need to Run DeepSeek Locally? VRAM by Version
DeepSeek currently has 7 versions you can deploy locally. The smallest, DeepSeek R1 Qwen3 8B, needs 8GB of VRAM at 4-bit; the largest, DeepSeek V4 Pro, is an MoE model and needs 16GB of VRAM plus about 837GB of RAM with MoE offloading. The cheapest device that gets one running is the Tesla V100 16GB, at about $370.
| Version | Released | Context | License | ||
|---|---|---|---|---|---|
| DeepSeek V4.1 FlashMoE | Oct 17, 2026 | 763B | 8 / 16B | 1M | MIT |
| DeepSeek V4 ProMoE | Aug 13, 2026 | 1,600B | 49B | 1M | MIT |
| DeepSeek V4 FlashMoE | Jul 31, 2026 | 284B | 13B | 1M | MIT |
| DeepSeek V3.2MoE | Dec 1, 2025 | 671B | 37B | 160K | MIT |
| DeepSeek R1 Qwen3 8BDense | May 29, 2025 | 8.2B | 8.2B | 128K | MIT |
| DeepSeek R1 Distill Qwen 32BDense | Jan 20, 2025 | 32.8B | 32.8B | 128K | MIT |
| DeepSeek R1 Distill Qwen 14BDense | Jan 20, 2025 | 14.8B | 14.8B | 128K | MIT |
About DeepSeek
DeepSeek is an AI company founded in Hangzhou in July 2023. Its founder, Liang Wenfeng, is also a co-founder of the quantitative hedge fund High-Flyer, and the company grew out of an AI research team that High-Flyer set up that April. Starting with the open-source releases of DeepSeek Coder and DeepSeek LLM in November 2023, it has released open weights for most of its main models at launch.
Two things made DeepSeek famous. The first is architecture. V2, in May 2024, introduced Multi-head Latent Attention (MLA), which saves VRAM, along with DeepSeekMoE, and V3 scaled up to 671B parameters that December. Later, V3.2 added DeepSeek Sparse Attention (DSA); V4, in April 2026, used compressed attention to reach a 1-million-token context; and V4.1 Flash, in September, switched to a causal encoder-decoder architecture with native image understanding. The second is reasoning models. After R1 came out in January 2025, the DeepSeek app briefly topped the US App Store's free chart, and that September the R1 paper was published in Nature after peer review. Since 2025, the weights of its main models have all used the MIT license, which allows commercial use.
- VendorDeepSeek
- Versions7
- Parameters8.2B – 1,600B
- Latest versionDeepSeek V4.1 Flash
- License
- MIT
- Official site
- www.deepseek.com
- Hugging Face org
- deepseek-ai
How much VRAM do DeepSeek models need at 2-bit?
2-bit, also called Q2, covers common GGUF files such as Q2_K and UD-Q2_K_XL. They are the smallest and lose the most quality, which makes them a way to get a model running first when VRAM is tight. Each row in the table below is a DeepSeek version and each column is a context length; each cell is the total VRAM needed to deploy it at Q2 (weights + KV cache + 1.1GB runtime overhead).
| Version | WeightsGB | 32K contextGB | 128K contextGB | 256K contextGB | 1M contextGB |
|---|---|---|---|---|---|
| DeepSeek V4.1 Flash763B (8 / 16B active) | 264.5GGUF | 265.6 | 265.7 | 265.8 | 266.5 |
| DeepSeek V4 Pro1,600B (49B active) | 638.4est. | 639.7 | 640.5 | 641.6 | 647.8 |
| DeepSeek V4 Flash284B (13B active) | 96.8GGUF | 98.1 | 98.6 | 99.3 | 103.7 |
| DeepSeek V3.2671B (37B active) | 246.6GGUF | 250 | 256.9 | — | — |
| DeepSeek R1 Qwen3 8B8.2B | 3.5GGUF | 9.4 | 23.9 | — | — |
| DeepSeek R1 Distill Qwen 32B32.8B | 13.1GGUF | 22.8 | 48.5 | — | — |
| DeepSeek R1 Distill Qwen 14B14.8B | 6.5GGUF | 14 | 33.3 | — | — |
How much VRAM do DeepSeek models need at 4-bit?
4-bit, also called Q4, corresponds to the widely used Q4_K_M files, which balance size and quality and are the default choice for local deployment. Each row in the table below is a DeepSeek version and each column is a context length; each cell is the total VRAM needed to deploy it at Q4.
| Version | WeightsGB | 32K contextGB | 128K contextGB | 256K contextGB | 1M contextGB |
|---|---|---|---|---|---|
| DeepSeek V4.1 Flash763B (8 / 16B active) | 444.7GGUF | 445.8 | 445.9 | 446 | 446.7 |
| DeepSeek V4 Pro1,600B (49B active) | 849.7GGUF | 851 | 851.8 | 852.9 | 859.1 |
| DeepSeek V4 Flash284B (13B active) | 155.1GGUF | 156.4 | 156.9 | 157.6 | 162 |
| DeepSeek V3.2671B (37B active) | 405.4GGUF | 408.8 | 415.7 | — | — |
| DeepSeek R1 Qwen3 8B8.2B | 5GGUF | 10.9 | 25.4 | — | — |
| DeepSeek R1 Distill Qwen 32B32.8B | 19.9GGUF | 29.6 | 55.3 | — | — |
| DeepSeek R1 Distill Qwen 14B14.8B | 9GGUF | 16.5 | 35.8 | — | — |
How much VRAM do DeepSeek models need at 8-bit?
8-bit, also called Q8, corresponds to Q8_0 files, which come closest to the original in quality and are also the largest. Each row in the table below is a DeepSeek version and each column is a context length; each cell is the total VRAM needed to deploy it at Q8.
| Version | WeightsGB | 32K contextGB | 128K contextGB | 256K contextGB | 1M contextGB |
|---|---|---|---|---|---|
| DeepSeek V4.1 Flash763B (8 / 16B active) | 508GGUF | 509.1 | 509.1 | 509.3 | 510 |
| DeepSeek V4 Pro1,600B (49B active) | 873.4GGUF | 874.7 | 875.5 | 876.6 | 882.8 |
| DeepSeek V4 Flash284B (13B active) | 161.9GGUF | 163.2 | 163.7 | 164.4 | 168.8 |
| DeepSeek V3.2671B (37B active) | 713.3GGUF | 716.7 | 723.6 | — | — |
| DeepSeek R1 Qwen3 8B8.2B | 8.7GGUF | 14.6 | 29.1 | — | — |
| DeepSeek R1 Distill Qwen 32B32.8B | 34.8GGUF | 44.5 | 70.2 | — | — |
| DeepSeek R1 Distill Qwen 14B14.8B | 15.7GGUF | 23.2 | 42.5 | — | — |
FAQ
Is DeepSeek open source and free for commercial use?
Yes. All 7 versions listed on this site are MIT-licensed, from the small R1-series models up to V4 Pro. You can use the weights and code commercially, modify them and redistribute them; the only condition is that copies keep the copyright and license notices. There is no monthly-active-user threshold and no add-on usage policy, and the R1-series model cards explicitly say that commercial use and distillation are supported. One thing to note: R1 Distill Qwen 14B and 32B were made by fine-tuning Qwen2.5 on data generated by R1, and their model cards point out that the base models were originally Apache 2.0-licensed. The base of R1 Qwen3 8B is Qwen3 8B Base.
Can you run DeepSeek with Ollama?
Only the small R1-series models run locally. deepseek-r1:8b now points to R1 Qwen3 8B (the 0528 release), and deepseek-r1:14b and deepseek-r1:32b are R1 Distill Qwen 14B and 32B. None of them is the full 671B R1. As for the large models, the Ollama pages for V3.2 and V4 Flash were taken down in July and September 2026, respectively. V4 Pro and V4.1 Flash have only :cloud tags, which run on Ollama's servers instead of your own GPU. To run V3.2, V4 Flash or V4 Pro locally, use mainline llama.cpp with unsloth's GGUFs (supported since May 2026 for V3.2 and since June for V4). llama.cpp support for V4.1 Flash hasn't been merged yet, so for now the only option is running the official weights on vLLM.
