DeepSeek

DeepSeek is a Chinese AI research company, backed by the quantitative hedge fund High-Flyer, that builds large language models and releases most of them as open weights under the MIT license. Its models compete with proprietary frontier systems at much lower prices, and its consumer chat app and API are used worldwide.

DeepSeek became widely known in early 2025. DeepSeek-V3, released in December 2024, used a mixture-of-experts (MoE) design with 671 billion total parameters and 37 billion active per token. DeepSeek-R1, a reasoning model trained from the V3 base and released on January 20, 2025, matched leading closed reasoning models on many benchmarks while publishing its weights and training method. The release prompted a sharp sell-off in AI hardware stocks and a broad debate about how much compute frontier models really need. Later releases followed quickly: V3.1 in August 2025, V3.2-Exp in September 2025 (which introduced DeepSeek Sparse Attention and cut API prices by more than half), and V3.2 in December 2025.

The current generation is DeepSeek-V4. On April 24, 2026 DeepSeek open-sourced a V4 preview in two sizes: V4-Pro (1.6 trillion total parameters, 49 billion active) and V4-Flash (284 billion total, 13 billion active), both with a 1-million-token context window and switchable thinking and non-thinking modes. The old deepseek-chat and deepseek-reasoner API names were retired on July 24, 2026. On September 10, 2026 DeepSeek released DeepSeek-V4.1-Flash, a 552-billion-parameter MoE model with a new causal encoder-decoder architecture that activates about 8 billion parameters when reading input and 16 billion when generating, plus native image understanding. DeepSeek says V4.1-Flash outperforms V4-Pro, so since September 14, 2026 API calls to V4-Pro are routed to V4.1-Flash until a V4.1-Pro ships. DeepSeek has not published a knowledge cutoff for the V4 family; its R1 paper gave July 2024 for the V3 base.

Open weights change how DeepSeek spreads. Companies self-host the models for data-residency reasons, cloud platforms serve them as cheap alternatives, and developers fine-tune them into specialist tools; the V4.1-Flash model page on Hugging Face listed 25 fine-tunes and 129 quantized versions in October 2026. A brand's representation in DeepSeek's parametric knowledge therefore travels into products that never call DeepSeek's own API. DeepSeek also runs a training crawler, DeepSeekBot, which you can allow or block in robots.txt.

Adoption is uneven by market. Italy's data protection authority had the app pulled from Italian app stores in January 2025, South Korea's privacy regulator saw new downloads suspended from February to April 2025, and Australia, Taiwan and the Czech government barred its use on public-sector devices. Those actions targeted the hosted app and service; self-hosted copies of the weights, like other open-source LLMs, keep data on the operator's own infrastructure.

For GEO, DeepSeek is worth tracking if your audience includes Chinese-speaking users, developer tools or companies that self-host models. Because many deployments run the model without live search, what DeepSeek says about your brand often comes from training data, which favors long-standing, widely repeated facts over last month's launch. Check its answers for outdated pricing, product names or positioning alongside your cross-platform AI visibility tracking.