DeepSeek V3
DeepSeekLLMs
90.0
Performance
★ 4.5
Rating
620
Reviews
Text GenerationReasoningOpen WeightLarge
About
Open-weight MoE model with 671B total parameters (37B active) delivering frontier performance at exceptionally low cost.
Strengths
Exceptional cost-efficiency: competitive with GPT-4o on most benchmarks at ~10x lower cost. Strong on math, coding, and Chinese language tasks. Open-weight with MoE architecture for efficient inference. Particularly good for production deployments where cost matters.
Specifications
- Context window
- 128,000
- Parameters
- 671B
Pricing
- Input cost
- $0.27 / 1M
- Output cost
- $1.10 / 1M
$0.27 input / $1.10 output per 1M tokens. Open-weight for self-hosting.
Speed & Latency
- 60
- tokens/sec
- 400ms
- time to first token
- 500 RPM
- rate limits
Available On
DeepSeek APIHuggingFaceTogether AIFireworks AI
Features
function callingstreamingjson modesystem messagesfill in the middle
Performance Trend
Benchmark score trends over time for the top 5 benchmarks.
Loading history...
Benchmarks
Scores from various benchmark tests; higher is better.
| Test | Score | Percentile | Source |
|---|---|---|---|
| BigBench Hard | 85.5 | p96 | seed |
| DROP | 85.0 | p95 | seed |
| GSM8K | 93.0 | p97 | seed |
| HumanEval | 89.2 | p97 |
