This model is 20 months old. A newer version of it has since shipped: DeepSeek V4 Flash Vision Exp (2026-08-21). The details below are still accurate for this model — it just is not what we would recommend today.
DeepSeek V3
DeepSeek🇨🇳 China
Context window128K tokens
Input / 1M tokens$0.27
Output / 1M tokens$1.1
Version History
DeepSeek-V3-1217minor
DeepSeek V3 December update with improved instruction following and expanded Chinese-English code-switching performance.
DeepSeek-V3major
671B MoE model released open-weights. Outperforms GPT-4o and Claude Sonnet 3.5.
Benchmark Scores
Full leaderboard →80.0%
AIME 2025
1358.0 elo
Arena Elo
89.4%
DocVQA
59.1%
GPQA
6.1%
Hallucination (Summarization)
90.0%
Hallucination Rate
93.5%
HumanEval
64.8%
LiveCodeBench
84.0%
MATH
88.5%
MMLU
75.2%
MMLU-Pro
72.5%
MMMU
70.0 tokens_per_sec
Speed (tok/s)
42.0%
SWE-bench Verified