CanItRun Logocanitrun.

DeepSeek R1 671B

DeepSeek R1 671B needs roughly 458.3 GB VRAM at Q4_K_M quantization (1503.6 GB at FP16). 3 GPUs we track can run it fully in VRAM at 8k context.

3 GPUs run this natively · 0 with CPU offload

DeepSeek671B params37B active (MoE)125k contextMITCommercial use ok

DeepSeek R1 671B is a Mixture of Experts (MoE) model with 671B total parameters but only 37B active per token developed by DeepSeek. January 2025 reasoning model trained via pure reinforcement learning (GRPO) without supervised fine-tuning. 671B MoE with 37B active parameters.

To run DeepSeek R1 671B locally: Full model needs datacenter hardware. Use distill variants: 32B distill (~20GB Q4) runs on 24GB GPU; 70B distill (~40GB Q4) needs dual GPU or Mac Studio. As a MoE model, inference speed depends on active parameters (37B) rather than total size.

AIME 2024: 79.8% pass@1, MATH-500: 97.3%, Codeforces: 96.3rd percentile — self-reflection and verification emerge from RL.

VRAM at each quantization

Figures below assume 8k context; KV cache grows linearly as context length increases.

QuantWeightsKV cacheTotal
FP322684.0 GB0.51 GB3006.7 GB
BF161342.0 GB0.51 GB1503.6 GB
FP161342.0 GB0.51 GB1503.6 GB
Q8_0713.3 GB0.51 GB799.4 GB
Q6_K550.9 GB0.51 GB617.6 GB
Q5_K_M477.8 GB0.51 GB535.7 GB
Q4_K_M408.6 GB0.51 GB458.3 GB
Q3_K_M322.8 GB0.51 GB362.1 GB
Q2_Krec255.7 GB0.51 GB286.9 GB
NVFP4cuda335.5 GB0.51 GB376.3 GB

KV cache is calculated at 8k context (FP16). Note that NVFP4 only runs on CUDA GPUs. Turn on TurboQuant in the calculator above for lower KV cache estimates.

Benchmarks

MMLU-Pro
85
GPQA Diamond
71.5
IFEval
83.3
MATH
97.3
LiveCodeBench
65.9

GPUs that run DeepSeek R1 671B natively (3)

Notes

Full reasoning model — outputs long chains-of-thought. Same architecture as V3; needs a multi-GPU server.

Hugging Face ↗Ollama ↗Released 2025-01-20

Compare DeepSeek R1 671B with other models

How to run DeepSeek R1 671B locally

816244880160320

Q2_K needs 286.9 GB — needs multiple datacenter-class GPUs (80 GB+ each).

Ollama

ollama run deepseek-r1:671b

llama.cpp

./llama-cli -m deepseek-r1.Q2_K.gguf -c 8192 -ngl 99

LM Studio: DeepSeek R1 671B is a massive reasoning MoE. Most users should look at the distilled variants (8B, 32B, 70B) instead. Only attempt the full model with a multi-GPU server.

Why this quantization? Same architecture as DeepSeek V3 (671B MoE, 37B active, MLA attention), but trained with reinforcement learning for chain-of-thought reasoning. Q2 is the only viable quantization for local deployment, requiring roughly 250 GB of VRAM. The reasoning capability is less sensitive to quantization degradation because the model's strength comes from its trained thinking patterns rather than raw parameter precision.

Who is DeepSeek R1 671B for?

Research labs and enterprises with server-class hardware who need the strongest open-weight reasoning model available. This is the full-scale version for teams with the infrastructure to run it -- most individual users should choose one of the distilled variants (8B, 32B, or 70B) instead.

Best for

  • Complex mathematical proofs and olympiad-level problem solving (MATH: 97.3)
  • Graduate-level science reasoning (GPQA: 71.5)
  • Competitive programming and algorithmic challenges (LiveCodeBench: 65.9)
  • Research tasks requiring explicit chain-of-thought reasoning traces
  • Benchmarking and evaluation as a frontier reasoning baseline

Not ideal for

  • Individual users without multi-GPU servers -- use the distilled 32B or 70B variants instead
  • Quick Q&A or chat -- the chain-of-thought output adds significant latency
  • Tasks where concise answers are preferred -- the model produces verbose reasoning traces by design

Continue reading

vram-guides10 min
DeepSeek Family Guide: R1, V3 & Distilled Models
hardware10 min
Used Enterprise GPUs for LLMs: A6000, A100, and Beyond

Frequently asked questions

What are the VRAM requirements for DeepSeek R1 671B?
DeepSeek R1 671B requires approximately 458.2 GB of VRAM at Q4_K_M quantization, 799.4 GB at Q8, and 1503.6 GB at FP16. These numbers assume 8k context window; VRAM scales linearly with context length due to the KV cache.
How many parameters does DeepSeek R1 671B have?
DeepSeek R1 671B has 671 billion total parameters, but only 37 billion are active per token thanks to its Mixture of Experts (MoE) architecture. This makes inference significantly faster than the total parameter count suggests.
Is DeepSeek R1 671B good at reasoning and math?
Yes. With a MATH score of 97.3 and MMLU-Pro of 85, DeepSeek R1 671B handles complex multi-step reasoning, analytical tasks, and problem-solving well.
Can DeepSeek R1 671B run on a 16 GB GPU?
No. At Q4_K_M, DeepSeek R1 671B needs 458.2 GB of VRAM — more than 16 GB. You will need a multi-GPU server.
Can DeepSeek R1 671B run on a 24 GB GPU?
No. Even at Q4_K_M, DeepSeek R1 671B needs 458.2 GB. Consider a multi-GPU server with 80 GB+ total VRAM.
What is the smallest quantization for DeepSeek R1 671B that fits in 24 GB of VRAM?
DeepSeek R1 671B cannot fit in 24 GB of VRAM at any standard quantization level. The minimum needed is 286.9 GB at Q2_K.
What GPU do I need to run DeepSeek R1 671B locally?
You need a multi-GPU server. At Q4_K_M, DeepSeek R1 671B needs 458.2 GB VRAM, more than any single consumer GPU. Consider 2–4× H100 or A100 GPUs.