DeepSeek: open-weight models and an AI assistant, with V3 and R1 on two tracks

DeepSeek is an open-source model family and AI assistant developed by DeepSeek (backed by High-Flyer), aimed at the base models and techniques behind general artificial intelligence. It means two things for a user: an assistant you can open in a browser and talk to, and open weights you can download and run yourself.

There are two main tracks on the model side. The general track is represented by DeepSeek-V3, built on a mixture-of-experts architecture with 671B total parameters and 37B activated per token, a 128K context window and generation speeds up to 60 TPS. The later V3.2 adds DSA fine-grained sparse attention to the architecture, using a lightning indexer to pick key information and greatly improving efficiency in long-text training and inference. The reasoning track is represented by DeepSeek-R1, trained from V3-Base and optimised with reinforcement learning, standing out in mathematics, programming and natural-language reasoning.

Branches have grown from those two tracks. R1-Zero skips supervised fine-tuning and relies purely on reinforcement learning, which brings strong reasoning but challenges in readability. R1-Distill distils reasoning data generated by R1 into small models ranging from 1.5B to 70B. R1-0528 is trained from V3-0324 with 660B parameters, emphasising deep reasoning, improved text generation and single-task runs lasting 30 to 60 minutes. On the multimodal side there is DeepSeek-VL2 (Tiny, Small and standard versions with 1.0B, 2.8B and 4.5B activated parameters) plus the Janus series, and vertically there is Prover-V2 for mathematical theorem proving.

In everyday use it is a broad assistant: question answering and multi-turn dialogue, text creation, translation, data processing, chart generation, code generation and debugging, mathematical reasoning, with web search and a deep-thinking mode on top. An image mode lets you upload a picture for visual understanding. For developers there is also a model development platform covering training, management and dataset control.

Main features

Question answering and dialogue

Handles questions on science, history, daily life and technology, follows context and stays coherent, which suits continuous follow-up questioning.

Writing and translation

Produces articles, stories, poetry, reports and emails, and translates between languages; DeepSeek-V3 covers as many as 83 languages.

Code generation and debugging

Generates code in many programming languages from natural-language descriptions and helps locate problems and improve implementations.

Mathematics and logical reasoning

Strong on maths and reasoning tasks, with a deep-thinking mode for questions that need multi-step analysis.

Web search and deep thinking

Search pulls in current information; deep-thinking mode shows the reasoning process for complex logic and multi-step work.

Data processing and charts

Cleans, analyses and runs statistics on data, then turns results into bar, line or pie charts.

Image understanding

Upload a picture for visual understanding and interaction; on the model side VL2 and Janus combine vision with language.

Document reading

Upload a document to extract key information and summarise it, which suits long material.

Development platform

A model platform supports training, management and dataset control, making it easier to wire the capability into your own systems.

How to use it

1. Pick an entry point. The web version works in any browser; the app is in the app stores; a browser extension is available in the Chrome Web Store.
2. Match the mode to the task. Everyday questions and copy use chat; current information uses AI search; documents use file reading; complex reasoning turns on deep thinking.
3. State the question clearly. Avoid vague phrasing, use keywords to mark what you need, and split complex work into smaller questions that go deeper one at a time.
4. Work across turns. Continuing in the same conversation to ask again, request a rewrite or add constraints usually works better than one very long prompt.
5. Borrow from the prompt library. The official library covers 13 scenario types, including code rewriting, code explanation, code generation, content classification, structured output, role play, prose and poetry, copy outlines, slogans, prompt generation and Chinese-English translation.
6. Deploy locally when you need a private environment. Download the model files from the official site, install dependencies, configure a server and deploy, then test and tune performance.

Where it fits

• Developers and algorithm teams: self-host from open weights, or use the platform for training, dataset management and model control.
• Content and marketing: generate marketing copy, meeting notes and first drafts at volume, then converge them by hand.
• Data roles: turn natural-language questions into queries, clean and analyse data, and produce charts that explain a conclusion quickly.
• Education and research: teachers generate lesson plans and materials and give students tailored paths, while maths work can use Prover-V2 for formal proof checking.
• Organisations with data compliance rules: put the model inside their own environment through local deployment so material does not leave.

Limits and things to keep in mind

• Deep thinking is slower. Showing the reasoning means more computation and waiting, so it is unnecessary for simple tasks; the material notes that R1-0528 can handle single tasks lasting 30 to 60 minutes, which suits hard problems rather than routine questions.
• Output still needs checking. Models produce plausible text that can be factually wrong, so figures, citations and conclusions should be verified against primary sources rather than treated as a source of fact.
• Open weights are not plug-and-play. Local deployment needs dependencies, a server and enough memory, plus performance tuning; running large parameter counts on small hardware is not realistic.
• Versions move fast. V3, V3.1-Terminus, V3.2, R1 and R1-0528 follow one another quickly with different capabilities and parameter counts, so choose against the current official documentation.
• Price and quotas come from the official source. The material here contains no confirmed pricing, rate limits or commercial terms, so check the official pages for cost and licensing instead of inferring from this article.

FAQ

What ways are there to use DeepSeek?

Web, app and browser extension, plus self-hosting of the open models; the development platform covers training and management.

How do V3 and R1 differ?

V3 is the general language model with mixture-of-experts, 671B total and 37B active parameters and a 128K context; R1 takes the reasoning-optimised route and uses reinforcement learning to improve maths, coding and natural-language reasoning.

How much text can it handle?

V3 supports a 128K context, and R1-0528 emphasises single tasks lasting 30 to 60 minutes; exact limits follow the official documentation.

Can it be deployed locally?

Yes. The path is to download the model files, install dependencies and the environment, configure a server, deploy, then test and tune performance, which suits scenarios with data-security requirements.