DeepSeek-V3.1

DeepSeek🇨🇳 China
deprecated

DeepSeek's first hybrid thinking/non-thinking model — one checkpoint serving both chat and reasoning via template switch. Faster than R1 at comparable quality, tuned for agentic tool use.

Context window131K tokens
Input / 1M tokens$0.56
Output / 1M tokens$1.68

Version History

deepseek-v3-1-launchmajor

DeepSeek-V3.1 launches. DeepSeek's first hybrid thinking/non-thinking model — one checkpoint serving both chat and reasoning via template switch. Faster than R1 at comparable quality, tuned for agentic tool use.

Benchmark Scores

Full leaderboard →
66.3%
AIME 2024
49.8%
AIME 2025
1418.0 elo
Arena Elo
74.9%
GPQA
5.5%
Hallucination Rate
56.4%
LiveCodeBench
83.7%
MMLU-Pro
66.0%
SWE-bench Verified