DeepSeek V3
Frontier-class quality at fast-tier prices — and open weights.
Intelligence index
67/ 100
vs all models41th pctile
Composite of MMLU, GPQA, MATH & HumanEval
Speed
90tok/s
vs all models45th pctile
Median across providers, steady state
Blended price
$0.48/ 1M tokens
vs all models68th pctile
3:1 input:output blend
At a glance
- Context window
- 64k tokens
- Max output
- 8k tokens
- Input price
- $0.27 / 1M tokens
- Output price
- $1.10 / 1M tokens
- Time to first token
- 0.45s
- Input modalities
- text
- Output modalities
- text
- License
- Open source
- Provider
- DeepSeek
Benchmark scores
Public scores from each provider; bars compare this model against the leader in each benchmark.
MMLU
General knowledge across 57 subjects
88.5
leader: 91.8
MMLU Pro
Harder MMLU successor with more reasoning
75.9
leader: 80.0
GPQA
Graduate-level science Q&A
59.1
leader: 78.0
MATH
Competition mathematics
84.1
leader: 94.8
HumanEval
Python code generation pass@1
89.5
leader: 95.8
Strengths
- Best $/intelligence on the market
- Open weights
Weaknesses
- Smaller context vs newer models
Best for
- Production chatbots
- Code assistance
- High-volume APIs
Models you should also evaluate
DeepSeek
DeepSeek R1
Open-weights reasoning model that matches o1 at 1/25 the price.
73 intel60 tok/s$0.96 /1M
Meta
Llama 3.3 70B
Open-weights 70B that matches GPT-4o on most benchmarks.
66 intel200 tok/s$0.27 /1M
Meta
Llama 3.1 405B
The original "open-source GPT-4" — largest publicly-released weights.
64 intel32 tok/s$2.70 /1M
DeepSeek V3 vs… popular head-to-heads
One-click matchups against the models people compare DeepSeek V3 with most.
DeepSeek V3 — frequently asked questions
DeepSeek V3 is a large language model from DeepSeek, released on 26 December 2024. Frontier-class quality at fast-tier prices — and open weights.
Need help choosing between models?
Compare every option in one sortable table — intelligence, speed and price on a single page.