deepseek-chat
by deepseek
Pricing
Input
$0.11 / 1M tokens
Output
$0.44 / 1M tokens
DeepSeek Chat (DeepSeek v3) Release by DeepSeek
Overview
DeepSeek Chat (DeepSeek v3) is a large-scale, mixture-of-experts AI model released in March 2025, featuring a massive 128K token context window and strong capabilities in reasoning, coding, and multilingual tasks.
- Release Date: March 24, 2025
- Architecture: Mixture-of-Experts (MoE), 671B parameters total, 37B active experts per token
- Context Length: 128,000 tokens
- Maximum Output Length: 8,000 tokens
- Inference Speed: ~60 tokens per second
- Latency: ~3.03 seconds to first token
- Training Data Cutoff: July 1, 2024
Features
- Enhanced logical reasoning and coding capabilities
- Multilingual support
- Advanced integration with external tools and APIs
- High throughput with MoE architecture
Benchmarks
| Task | DeepSeek Chat | Claude 3 Opus | Gemini 2.5 Flash |
|---|---|---|---|
| MMLU (Reasoning) | 79.2% | 86.8% | 77.9% |
| MGSM (Math) | 81.3% | 60.1% | 79.7% |
| HumanEval (Coding) | 76.5% | 84.9% | 74.1% |
| MMMU (Multimodal) | 57.3% | 54.9% | 56.8% |
Pricing
| Type | Price per 1M tokens |
|---|---|
| Input | $0.27 |
| Output | $1.10 |
Use Cases
- Real-time chatbots and customer support
- Educational tutoring and study assistance
- Content generation and summarization
- Software development and debugging
- Data analysis and insights
Safety and Stability
- Open source under MIT license
- No live internet connection (offline model)
- Strong instruction following and robustness
- Training data cutoff: July 1, 2024
Limitations
- Only text input/output supported (no multimodal yet)
- Possible numeric precision inconsistencies
- Latency higher compared to smaller models due to size
License
DeepSeek Chat (DeepSeek v3) is licensed under the MIT License and allows commercial use.