DeepSeek
A Chinese series of large language models and a free chatbot app
DeepSeek is a family of large language models built by the Chinese company DeepSeek, starting in November 2023. Its R1 release of January 2025 topped the US iOS free app chart within days, drawing attention to cheap, openly released models from China.
- R1 topped the US iOS free app chart within 17 days of the chatbot's launch
- Nvidia shares dropped 18% on 27 January 2025
- V2 was priced at 2 RMB per million output tokens
- DeepSeek-MoE activates 2.7B of its 16B parameters per token
- V3 was trained on 14.8T tokens of mostly English and Chinese text
What it is
DeepSeek is a line of large language models made by the Chinese firm DeepSeek, with work starting in November 2023. The family covers general chat models, code assistants, maths provers and a reasoning model. Sizes range from small 7B designs up to a 671B maths prover. A free chatbot app for iOS, Android and the web sits on top of these models and answers questions, handles logic puzzles and writes code at a level comparable to rival chatbots on the benchmarks US AI firms use.
How it came about
The DeepSeek-LLM line appeared in November 2023 in 7B and 67B sizes. V2 followed in May 2024, then the chatbot app on 10 January 2025, built on DeepSeek-R1. By 27 January that app had overtaken ChatGPT as the most downloaded free title on the US iOS store, and Nvidia's shares fell 18% the same day. A large-scale attack that day forced sign-ups to be restricted. Later came DeepSeek-GRM with Tsinghua University in April 2025, the maths prover Prover-V2-671B on 30 April 2025, and V4, V4-Pro and V4.1-Flash during 2026.
Inside the model
Early releases were dense decoder-only transformers close in shape to Llama: pre-norm layers with RMSNorm, SwiGLU feedforward blocks, rotary embeddings and grouped-query attention, a 102,400 byte-level BPE vocabulary and 4,096-token context. V2 swapped standard attention for multi-head latent attention, which compresses key-value caches to save memory, and added a mixture-of-experts feedforward layer where a gate picks the highest-scoring experts per token. A DeepSeekMoE variant keeps shared experts always active alongside routed ones, easing load balancing. V3 added multi-token prediction and heavy low-level tuning: 8-bit floats in most of the forward pass, custom 12-bit values after attention, and communication overlapped with compute across H800 clusters.
What it can do
Benchmark tests used by American AI companies put the chatbot on a par with other systems at answering questions, logic and coding. The DeepSeek-LLM paper claimed scores above Llama 2 and most open models of its day. DeepSeek-V2 placed seventh on the University of Waterloo Tiger Lab leaderboard, and the Financial Times noted its low price of 2 RMB per million output tokens. V3 was reported to beat Llama 3.1 and Qwen 2.5 while reaching the level of GPT-4o and Claude 3.5 Sonnet.
How to use it
The chatbot is free on the web and on iOS and Android, with no cap on question count, though the web version allows 500 messages an hour to block spam bots. Sign-up uses a global email service such as Gmail or Yahoo. R1 and V3 are also sold through an API priced per token: about $0.28 per million in and $0.42 per million out as of February 2025. Weights for R1, V3 and V4 carry the MIT License; earlier models use the DeepSeek License with downstream-use conditions.
Safety and testing
A large-scale cyberattack on 27 January 2025 disrupted the servers, after which new accounts were limited to mainland China phone numbers, email addresses or Google logins. V2 training included 0.3M safety examples and a reward model for helpfulness and safety built from human preference data, plus a hand-written rule model. Concerns have been raised about the chatbot's compliance with Chinese censorship rules and its data collection, drawing regulatory attention in several countries.
Reception and debate
The R1 app's rapid climb to the top of the US iOS chart, and the fall in Nvidia's share price, were read by commentators as a shake-up of the AI field and the start of a global race. Coverage has praised the open weights, released infrastructure code, energy efficiency and help to open-source AI, while faulting privacy and information control. Cost figures for V3 training have been called misleading because they cover only part of the real outlay.
Why it matters
By publishing weights and training details for models that compete with far larger labs, DeepSeek is credited with shifting expectations about what a smaller team can achieve cheaply. Its mixture-of-experts variant with shared experts, the latent attention design and its reinforcement learning recipe have fed into later work, and the company has said newer GRM models will also be released openly.
Questions people ask about DeepSeek
What is DeepSeek?
It is a family of large language models from the Chinese company DeepSeek, first released in November 2023, plus a free chatbot app built on them. The range spans general chat, code, maths and reasoning models of many sizes.
How can I use it?
The chatbot runs free on the web and on iOS and Android, with a 500 messages per hour limit on the web. Developers can also call R1 and V3 through a paid API, roughly $0.28 per million input tokens and $0.42 per million output tokens as of February 2025.
Are the models open?
R1, V3 and V4 weights are released under the MIT License. Earlier models such as DeepSeek-LLM and DeepSeek Coder use the DeepSeek License, which allows use but adds conditions on downstream behaviour. The apps themselves are proprietary.
Why did it attract so much attention in January 2025?
Its R1-based app overtook ChatGPT as the most downloaded free app on the US iOS store by 27 January 2025, and Nvidia's shares fell 18% that day. Observers described it as a shift in the AI race toward cheaper, openly published models.
What concerns have been raised?
Critics point to its compliance with Chinese censorship policies and its data handling, which have drawn regulatory scrutiny in several countries. A large-scale cyberattack in January 2025 also forced limits on new sign-ups.
Good reading and sources
- Official websiteOfficial site
- Why DeepSeek Could Change What Silicon Valley Believes About A.I. (Published 2025)The New York Times, Article
- China's DeepSeek AI dethrones ChatGPT on App Store: Here's what you should knowCNBC, Article
- ‘Sputnik moment’: $1tn wiped off US stocks after Chinese firm unveils AI chatbotThe Guardian, Article
- DeepSeek (AI) on WikipediaEncyclopedia
- More photos on Wikimedia CommonsPhotos
- Wikidata recordOpen data
The facts on this page come from the Wikipedia article "DeepSeek (AI)", written by its contributors and used under the CC BY-SA 4.0 licence, and from its Wikidata record (CC0). EveryLLM rewrote the text in its own words, with AI assistance, and compares every page with the article to make sure no passage is copied. It may be reused under the same CC BY-SA 4.0 licence with a link back.
It may contain mistakes, and figures change over time. Nothing here is investment or buying advice. If something is wrong, tell us. Last refreshed 2026-11-17.