LLMOpen weights
DeepSeek-V4-Pro
DeepSeek
Updated May 21, 2026 · Benchmarks via Artificial Analysis, specs & pricing via OpenRouter · Methodology · Report an error
Specifications
- Type
- LLM
- Access
- Open weights
- Released
- April 24, 2026
- License
- MIT
- Parameters
- 1.6T (49B active)
- Context window
- 1,048,576 tokens
- Output speed
- 30 tok/s
- Latency (TTFT)
- 1.16s
- Input
- Text
- Output
- Text
- API pricing
- $0.44 in · $0.87 out / 1M tokens
Capabilities
Function callingStructured output
Benchmarks
Reasoning
| GPQA Diamond | 90.1 |
Coding
| LiveCodeBench | 93.5 |
| SWE-bench Verified | 80.6 |
General
| MMLU-Pro | 87.5 |
| Humanity’s Last Exam | 35.9 |
Our take
DeepSeek’s frontier 1.6T-parameter MoE with new hybrid attention and a native 1M context, closing much of the gap to top closed models among open weights.