Official links | Chat | Code Plan | Agent tools | API |
|---|---|---|---|---|
| Global | — | — | ||
| Mainland China | — | — |
English reports linked to this entity will appear here as their translations are published.
DeepSeek official documentation confirms that deepseek-v4-pro now serves DeepSeek-V4-Pro-0813 under the unchanged API model name, with a one-million-token context window, up to 384K output tokens, and Responses API support.
DeepSeek releases the V4 Flash 0731 API release, with post-training that substantially improves agent capabilities and native support for the Responses API and Codex.
DeepSeek launches V4-Pro and V4-Flash in its API and begins replacing the older chat and reasoner model names.
Another fine-tuning iteration on the V3 base, focused on lowering peak memory use and speeding up inference on local hardware.
An architecture tune of V3 that improves MoE expert load-balancing and the stability of very large-scale distributed training.
DeepSeek passes ChatGPT on the U.S. App Store's free chart as Nvidia loses roughly $593 billion in market value during an AI-sector selloff.
DeepSeek publishes R1, R1-Zero, six distilled models, and its reinforcement-learning recipe, pushing open reasoning models into the frontier conversation.
DeepSeek puts V3, web search, a reasoning mode, and file analysis into a free mobile app five days before the R1 release.
DeepSeek releases a 671B MoE model with 37B active parameters, open weights, low API prices, and unusually detailed efficiency disclosures.
R1-Lite-Preview exposes long-form reasoning on the web and promises an open model and API, showing that R1 was already underway before V3's release.
DeepSeek releases a V2 architecture model optimized for programming and mathematical reasoning.
DeepSeek releases a 236B mixture-of-experts model that combines sparse activation with latent attention to cut training and inference costs.
DeepSeek releases math models and Group Relative Policy Optimization, creating a direct algorithmic predecessor to R1's reinforcement-learning recipe.
DeepSeekMoE uses fine-grained experts and shared expert isolation to improve sparse-model efficiency, with 16B weights and training code released.
DeepSeek releases code models from 1.3B to 33B parameters, making open weights and code part of its product strategy from the outset.
