DeepSeek V3 is the 671 B MoE base model (37 B active) that DeepSeek used as the starting point for R1. It beats Llama 3.1 405B on most benchmarks at a fraction of the compute, thanks to MLA attention and FP8 mixed-precision training. Inference VRAM mirrors R1 — ~370 GB at Q4_K_M, requiring multi-GPU datacenter hardware. Recommended over R1 when you want fast responses without the <think> overhead.
DeepSeek V3 671B needs at minimum 406 GB of VRAM at Q4_K_M quantization (1476.2 GB at FP16). The cheapest GPU that comfortably fits with KV-cache headroom is the NVIDIA GH200 Grace Hopper (576 GB VRAM, $65,000 MSRP). Community benchmark submissions are open. This model exceeds 48 GB at Q4, so plan for a 2-or-more-GPU split.
Weights only. Add ~20-30% for KV cache at typical context lengths.
| Quant | VRAM | Quality | Notes |
|---|---|---|---|
| FP16 | 1476.2 GB | Reference | Training-precision weights |
| Q8_0 | 738.1 GB | Near-lossless | 8-bit, ~0.1% perplexity hit |
| Q6_K | 597.9 GB | Very high | 6-bit, near Q8 quality |
| Q5_K_M | 509.3 GB | High | Strong middle ground |
| Q4_K_M | 406.0 GB | Balanced (recommended) | Default for local deployments |
| Q4_0 | 415.2 GB | Legacy 4-bit | Older GGUF, kept for compatibility |
| Q3_K_M | 317.4 GB | Lossy | When VRAM is very tight |
| Q2_K | 236.2 GB | Extreme | Rescue option, quality degrades visibly |
Filtered from MyAIHardware's GPU database. FP16 needs 1476.2 GB; Q4_K_M needs 406.0 GB.
| GPU | Per-card VRAM | 2x | 4x | 8x |
|---|---|---|---|---|
| AMD Instinct MI355X [VERIFY] | 288 GB | 576 | 1152 | 2304 |
| AMD Instinct MI325X | 256 GB | 512 | 1024 | 2048 |
| Apple Mac Studio (M3 Ultra, 256 GB) | 256 GB | 512 | 1024 | 2048 |
| AMD Instinct MI300X | 192 GB | 384 | 768 | 1536 |
| Apple Mac Studio (M2 Ultra, 192 GB) | 192 GB | 384 | 768 | 1536 |
| Apple Mac Pro (M2 Ultra, 192 GB) | 192 GB | 384 | 768 | 1536 |
We don't have community benchmarks for this exact model yet.
Official weights + popular runtime tags.
ollama pull deepseek-v3Step-by-step guides that use this model.
Models in a similar size or capability class.
Use these in papers, blog posts, or internal docs.
@misc{deepseek-v3-671b-2024,
title={DeepSeek V3 671B},
author={DeepSeek AI},
year={2024},
url={https://huggingface.co/deepseek-ai/DeepSeek-V3}
}DeepSeek AI (2024). DeepSeek V3 671B [Model card]. Hugging Face. https://huggingface.co/deepseek-ai/DeepSeek-V3
DeepSeek V3 671B (DeepSeek V3, DeepSeek AI, 2024). Available at https://huggingface.co/deepseek-ai/DeepSeek-V3.