DeepSeek

DeepSeek V3 671B

Name: DeepSeek V3 671B
Author: DeepSeek

Aktuell

HuggingFace

Ollama

1.1MDownloads4.1KLikesDec 2024Veröffentlicht131K TokenKontextDeepSeekLizenz75 StarkQualität

DeepSeek V3 671B (671B parameters) requires approximately 470.4 GB of VRAM with Q4_K_M quantization. As a Mixture of Experts model with 37B active parameters, it uses less memory than its total parameter count suggests. For the best balance of quality and speed, we recommend hardware with at least 541 GB of VRAM.

Loslegen

— kopieren & einfügen, um lokal auszuführen

Copy-paste commands to run DeepSeek V3 671B on your machine.

Run

ollama run deepseek-v3

Quick specs

Parameters671B (37B active)

Architecturemoe (MoE)

Context131K tokens

Modalitytext

Min RAM261.7 GB

Rec. RAM409.3 GB (Q4_K_M)

LicenseDeepSeek

FamilyDeepSeek

✓ Chat✓ Reasoning

About this model

We present DeepSeek-V3, a strong Mixture-of-Experts (MoE) language model with 671B total parameters with 37B activated for each token. To achieve efficient inference and cost-effective training, DeepSeek-V3 adopts Multi-head Latent Attention (MLA) and DeepSeekMoE architectures, which were thoroughly validated in DeepSeek-V2. Furthermore, DeepSeek-V3 pioneers an auxiliary-loss-free strategy for load balancing and sets a multi-token prediction training objective for stronger performance.

•On top of the efficient architecture of DeepSeek-V2, we pioneer an auxiliary-loss-free strategy for load balancing, which minimizes the...
•We investigate a Multi-Token Prediction (MTP) objective and prove it beneficial to model performance

Verwandte Modelle

Quantisierungsoptionen

VRAM-Schätzungen nach Quantisierungsstufe

No hardware detected — fit column shows raw VRAM estimates

Quant	Bits	VRAM	Quality	Fit
Q2_K	2	261.7 GB	Low	—
Q3_K_S	3	328.8 GB	Low	—
NVFP4	4	375.8 GB	Medium	—
Q4_K_M	4	409.3 GB	Medium	—
Q5_K_M	5	483.1 GB	High	—
Q6_K	6	550.2 GB	High	—
Q8_0	8	718.0 GB	Very High	—
F16	16	1375.6 GB	Maximum	—

Quality benchmarks

DeepSeek V3 671B benchmark scores

Benchmark verified

Coding

SWE-bench Verified42.0%

HumanEval+82.6%

Aider Polyglot49.6%

LiveCodeBench37.6%

Reasoning

MMLU-Pro75.9%

GPQA Diamond59.1%

MATH-50090.2%

ARC Challenge—

General

Chatbot Arena—

IFEval86.1%

Source: official · 2024-12-26

Hardware-Kompatibilität

Eignungsschätzungen für alle Hardware

Rechner öffnen

Computing compatibility...

Speicheraufschlüsselung

Reference: RTX 2060 6GB

Weights409.3 GB

KV Cache59.6 GB

Runtime0.9 GB

Headroom0.6 GB

Häufig gestellte Fragen

FAQ — DeepSeek V3 671B

Siehe auch

Quantisierungsleitfaden Bewertungsmethodik Rechner öffnen