What Do You Need to Run DeepSeek Locally? VRAM by Version
DeepSeek currently has 5 versions you can deploy locally. The smallest, DeepSeek R1 Qwen3 8B, needs 7GB of VRAM at Q4; the largest, DeepSeek V3.2, needs 12GB. The cheapest device that gets one running is the GeForce RTX 2080 Ti, with no used-price data yet.
VRAM needed for each DeepSeek version
5 versions| Version | Released | Context | ||||
|---|---|---|---|---|---|---|
| DeepSeek R1 Qwen3 8BDense | 8.2B | May 29, 2025 | 128K | 7 | GeForce RTX 2080 Ti— · 11 GB | 1,003,991 |
| DeepSeek R1 Distill Qwen 14BDense | 14.8B | Jan 20, 2025 | 128K | 11 | GeForce RTX 2080 Ti— · 11 GB | 396,404 |
| DeepSeek R1 Distill Qwen 32BDense | 32.8B | Jan 20, 2025 | 128K | 22 | Arc Pro B60— · 24 GB | 562,818 |
| DeepSeek V4 FlashMoE | 284B (13B active) | Jul 31, 2026 | 1M | 7plus 155 GB of RAM | Mac Studio M3 Ultra 512GB— · 512 GB | 4,559,659 |
| DeepSeek V3.2MoE | 671B (37B active) | Dec 1, 2025 | 160K | 12plus 365 GB of RAM | Mac Studio M3 Ultra 512GB— · 512 GB | 1,559,697 |
Minimum VRAM = Q4_K_M weights + KV cache for an 8K context + 1GB runtime overhead, rounded up. The cheapest device is the first one that fits when sorting by median used price on the current price basis.
About DeepSeek
DeepSeek is published by DeepSeek under MIT. This page covers 5 versions ranging from 8.2B to 671B parameters; structure parameters and quantized sizes are synced weekly from Hugging Face.
VendorDeepSeek
LicenseMIT
Official sitehttps://www.deepseek.com/
Hugging Face orgdeepseek-ai