Major Release DeepSeek LLM DeepSeek License (open weights)

DeepSeek DeepSeek-V3

DeepSeek's breakthrough open-weight MoE rivaling GPT-4-class quality.

Released 2024-12-26 · 671B · 128K tokens context · knowledge cutoff 2024-07

Overview

DeepSeek's most advanced open-source model with MoE architecture

Specifications

DeveloperDeepSeek
Release date2024-12-26
Model typeLLM
Parameters671B
ArchitectureSparse MoE (37B active / 671B total)
Context window128K tokens
Max output8K
Knowledge cutoff2024-07
LicenseDeepSeek License (open weights)
Input modalitiestext
Output modalitiestext

Benchmarks

Benchmark DeepSeek-V3
MMLU 88.5%
HumanEval 82.6%
MATH 61.6%

Pricing

$0.27
per 1M input tokens
$1.10
per 1M output tokens

Availability

Open Source DeepSeek API Hugging Face Together AI

Official references