Last updated:
DeepSeek
The Chinese open-weight model family that shook up the industry by matching closed frontier models at a fraction of the reported training cost.
Model Facts
Latest model: DeepSeek-V4-Pro 0813
- Provider
- DeepSeek
- Released
- Aug 13, 2026
- Context window
- 1M tokens
- Weights
- Open
Benchmarks
Full Model List
| Model | Released | Status | Note |
|---|---|---|---|
| DeepSeek-V4-Pro 0813 | Aug 13, 2026 | – | updated build: 1.6T parameters (49B active), MIT license |
| DeepSeek-V4 / V4-Pro | Apr 24, 2026 | – | V4 generation launch |
| DeepSeek-V4-Flash | Apr 24, 2026 | – | smaller sibling of V4-Pro, same release |
| DeepSeek-R1 | Jan 20, 2025 | – | visible step-by-step reasoning, made global headlines |
| DeepSeek-V3 | Dec 26, 2024 | – | matched leading closed models on several benchmarks |
| DeepSeek-V2 | May 6, 2024 | – | early Mixture-of-Experts release |
| DeepSeek-Coder | Nov 2, 2023 | legacy | first code-specialized DeepSeek models |
- Official model page:
- deepseek.com
- Official documentation:
- api-docs.deepseek.com
- Other sources:
- V4 Pro 0813 on Artificial Analysis
What it is
DeepSeek is the Chinese lab that shook up the industry by matching closed frontier models at a fraction of the reported training cost. The current generation is DeepSeek V4 (April 24, 2026), most recently refreshed as V4-Pro 0813 on August 13, 2026.
Specs
V4-Pro 0813 is a sparse mixture-of-experts reasoning model with 1.6 trillion total parameters and 49B active per token, a 1M-token context window, and text-only input and output. Weights are published on Hugging Face under the MIT license. V4-Flash is a smaller sibling released with V4.
Where these numbers come from
The benchmark is Artificial Analysis's independent Intelligence Index, where V4 Pro 0813 scores 36 (ranked #9 of 117 models in its comparison set, against a median of 18). Other published scores (for example on SWE-bench Verified) vary by source and are not shown.
Notable traits
- Open weights under a permissive MIT license.
- Very low API prices relative to closed frontier models.
- Text-only: no native image, audio or video input.