deepseek
DeepSeek is an open-source family of large language models known for its excellent value, strong reasoning, and 1M long context. It offers multiple tiers—V3 / V4 Pro / V4 Flash—and supports disk-based context caching, where cache hits cost only about 26% of the input price, delivering a major cost advantage in long-context scenarios.
1M token context window; disk cache (cache_hit ~26% of cache_miss); visual understanding; Function Calling; JSON mode; FIM completion; compatible with the OpenAI Chat Completions interface; Batch API discounts.
High-value reasoning tasks; code generation and completion; long-document analysis; Chinese-language understanding; agent workflows; batch data processing; academic research and education.
No image generation; caching depends on hits (not guaranteed); high concurrency requires contacting sales; some enterprise features require separate enablement; FP8 quantized versions are cheaper but with slightly reduced accuracy.
| Model | Input / 1M | Output / 1M | Cache Read / 1M | Per 1M |
|---|---|---|---|---|
| 1M | ||||
| 1M | ||||
| - | 1M | |||
| 1M | ||||
| - | 1M | |||
| 1M |