DeepSeek V3
DeepSeek · DeepSeek License
Mixture of Experts: 671B total parameters, 37B active per token.
Massive MoE model rivaling GPT-4 class. Only ~37B parameters active per token despite 671B total. Requires multi-GPU or very large unified memory (128GB+ Apple Silicon at Q2/Q3). Not for casual home use. Included for completeness and to show what the high end looks like.
- Parameters
- 671B
- Architecture
- MoE (37B active)
- Context
- 65,536 tokens
- Released
- 2024-12-26
- Engines
- llama.cpp, vLLM, SGLang
Parameters
671B
VRAM
360 GB
Context
64K
Formats
4
GPUs
41
DeepSeek V3 (671B) requires 360 GB VRAM at recommended quality (FP16). On NVIDIA Grace Blackwell Ultra GB300, expect approximately 22 tok/s at Q4_K_M.
Source: OwnRig methodology
360 GB
FP16
335 GB
64K tokens
Chat
VRAM Requirements
| Quality | Quantization | VRAM | File Size |
|---|---|---|---|
| full | FP16 | 360 GB | 335 GB |
| efficient | Q4_K_M | 180 GB | 168 GB |
| compressed | Q3_K_M | 145 GB | 135 GB |
| compressed | Q2_K | 115 GB | 108 GB |
Context Length Impact
KV cache VRAM at FP16 quality. Longer context = more memory.
| Context | KV Cache | Total VRAM |
|---|---|---|
| 2K | 1 GB | 361 GBexceeds 24 GB |
| 4K | 2 GB | 362 GBexceeds 24 GB |
| 8K | 4.1 GB | 364.1 GBexceeds 24 GB |
| 16K | 8.2 GB | 368.2 GBexceeds 24 GB |
| 32K | 16.4 GB | 376.4 GBexceeds 24 GB |
| 64K | 32.8 GB | 392.8 GBexceeds 24 GB |
Compatible GPUs
41 devicesShowing 41 of 41 entries
Frequently Asked Questions
- How much VRAM does DeepSeek V3 need?
- DeepSeek V3 requires 360 GB VRAM at recommended quality (FP16). At lower quality settings, it can fit in as little as 115 GB.
- What is the best GPU for DeepSeek V3?
- The NVIDIA Grace Blackwell Ultra GB300 delivers the best performance for DeepSeek V3, achieving 22 tok/s at Q4_K_M with an good rating.
- Can I run DeepSeek V3 on an RTX 4060 Ti?
- DeepSeek V3 at Q2_K requires 360 GB VRAM, which exceeds the RTX 4060 Ti's 16 GB. Consider a lower quantization or a GPU with more VRAM.
- What quantization should I use for DeepSeek V3?
- For the best quality, use FP16 (360 GB VRAM). If your GPU has limited VRAM, Q2_K (115 GB) is the most efficient option with acceptable quality.
- Is DeepSeek V3 good for coding?
- DeepSeek V3 supports coding use cases. For the best coding experience, pair it with an embedding model for local RAG.
Data confidence: estimated. Source
VRAM requirements are calculated from model parameters and may vary by inference engine, context length, and batch size. Performance estimates are based on community benchmarks and should be verified for your specific configuration.DeepSeek is a trademark of its respective owner. OwnRig is not affiliated with or endorsed by the model creator.