# DeepSeek: reviews and analysis > The Chinese assistant running V4 with open weights, completely free. - Canonical: https://serchai.com/en/reviews/deepseek/ - Site: Serchai (https://serchai.com) — AI tools comparator - Language: en - Updated: 2026-07-26 --- ## Verdict DeepSeek delivers reasoning close to the paid assistants while charging nothing for the chat, and it publishes the V4 weights under an MIT license, so you can take it to your own server. You pay the bill in governance: its policy states data is processed in China, censorship is visible on sensitive topics, and several governments have banned the app on official devices. **Best for:** People who want serious reasoning with no monthly fee and who are not handling client data or regulated information. **Rating:** 4.1/5 ## Pros - The chat is entirely free, with no paid tier - Open V4 weights under an MIT license so you can self-host - API from $0.14 per million input tokens ## Cons - Its policy states data is processed and stored in China - Visible censorship on topics sensitive to the Chinese government - Minimal product: no memory, no connectors, no desktop apps ## Key facts - Price: Free - Free trial: No - Platforms: Web, iOS, Android - Categories: [Productivity](https://serchai.com/en/best-ai/productivity/) - Official website: https://chat.deepseek.com ## What the internet says (agentic sweep) The same argument shows up everywhere: near-frontier reasoning for nothing in the chat and for pennies on the API, with V4 weights published on Hugging Face under an MIT license. The objection repeats too and it is not a small one, DeepSeek's own policy states data is processed and stored in China, Italy's Garante ordered a limitation on processing Italian users' data, and several governments have banned the app on official devices. - Sweep date: 2026-07-26 - Derived score: 4.1/5 ### Axes - Results: 4.3/5 - Control: 4/5 - Real price: 4.8/5 - Integration: 3.6/5 - Support: 3.8/5 ### Recurring themes in favor - The chat is entirely free, with no paid tier and no paywall, and that is the most cited reason people keep it open as a second opinion (strong theme) - Open weights for V4-Pro and V4-Flash under an MIT license on Hugging Face, so anyone can download them and run them on their own infrastructure (strong theme) - Official API pricing starts at $0.14 per million input tokens on Flash, an order of magnitude below comparable Western providers (strong theme) - Long-form reasoning and coding hold up against paid assistants, especially with thinking mode switched on (present theme) - Its API documentation gets cited as more concise and to the point than what Western providers publish (present theme) ### Recurring themes against - Its privacy policy states it collects, processes and stores personal data in the People's Republic of China and uses it to train its models, with an opt-out available (strong theme) - Italy's Garante ordered a limitation on processing Italian users' data in January 2025, and Australia, South Korea, Taiwan, the Czech Republic and several US agencies banned the app on official devices (strong theme) - Visible censorship on topics sensitive to the Chinese government, where the assistant starts writing an answer then deletes it before finishing (strong theme) - Capacity limits on the larger model, with waits, timeouts and rejected requests at peak hours (present theme) - Minimal product around the model: no desktop apps, no memory across chats, no connectors and no image generation (present theme) ### Sweep sources - [Official pricing] https://api-docs.deepseek.com/quick_start/pricing — Tarifa oficial verificada: deepseek-v4-flash desde 0,14 $ por millón de tokens de entrada y 0,28 $ de salida, deepseek-v4-pro desde 0,435 $ y 0,87 $, contexto de 1M en ambos - [Docs] https://api-docs.deepseek.com/news/news260424/ — Anuncio oficial del 24 de abril de 2026: V4 Preview en dos variantes, contexto de 1M por defecto y retirada de deepseek-chat y deepseek-reasoner el 24 de julio de 2026 - [Docs] https://cdn.deepseek.com/policies/en-US/deepseek-privacy-policy.html — Política de privacidad oficial: los datos personales se recogen, procesan y almacenan en la República Popular China y se usan para entrenar sus modelos, con derecho de exclusión - [Communities] https://news.ycombinator.com/item?id=47884971 — Hilo del lanzamiento de V4 con más de 2.000 votos y 1.600 comentarios: elogios al precio, a los pesos abiertos y a la documentación, quejas por límites de peticiones y falta de visión - [App stores] https://apps.apple.com/us/app/deepseek-ai-assistant/id6737597349 — Ficha de la App Store: 4 sobre 5 con unas 11.000 valoraciones, gratis, sin plan de pago listado. Las reseñas repiten censura en temas de China y cambios de idioma - [Press] https://simonwillison.net/2026/Apr/24/deepseek-v4/ — Análisis del lanzamiento: V4 cerca de la frontera con un desfase que el propio informe técnico cifra en 3 a 6 meses, pesos MIT en Hugging Face - [Press] https://www.garanteprivacy.it/home/docweb/-/docweb-display/docweb/10097450 — Comunicado del Garante italiano del 30 de enero de 2025: limitación urgente del tratamiento de datos de usuarios italianos por falta de garantías en la transferencia a China > **TLDR:** DeepSeek is the assistant from the Chinese company of the same name, and its chat costs nothing at all, with no paid tier and no paywall. Since April 2026 it runs on the V4 family, which also ships as open weights under an MIT license. It gets close to the big paid assistants on reasoning and coding without a monthly bill. The cost sits elsewhere: its own policy states data is processed in China, and several governments have banned the app on official devices. ## What DeepSeek is and how it works DeepSeek is a conversational assistant from Chinese company DeepSeek AI, reachable from the browser and from its iOS and Android apps. Its current model is the V4 family, released as a preview on 24 April 2026 in two variants, V4-Pro for long reasoning and V4-Flash for fast answers. The R2 model many people treated as inevitable never shipped, so anyone looking for that name will find V4 instead. What separates DeepSeek from other generalist assistants is how it publishes. The weights for both models sit on Hugging Face under an MIT license, so anyone can download them, run them on their own hardware and modify them without asking permission. No major commercial assistant does that with its flagship model, and for a company worried about vendor lock-in that exit door is worth more than any feature. The chat offers a direct mode and a thinking mode, with a one-million-token context window across the official services. You can paste a full contract, a code repository or a thesis and keep asking about the same material without chopping it up by hand. ## What it's like to use day to day The interface is among the most bare in the category: a text box, a history list and two toggles for reasoning and web search. There are no project folders, no memory across conversations, no connectors to your mail or your drive, and no image generation. If you come from an assistant with an ecosystem around it, the feeling is of going back three years in product and forward a fair way in model. Where it delivers is on what you actually asked it for. Mathematical reasoning, long document analysis, coding and refactoring hold up against the paid assistants, especially with thinking mode on. People who have run both models agree that Flash performs above what its size suggests, and that Pro lands a few months short of the frontier, a gap the company's own technical report admits. Two frictions are real. The first is capacity, since at peak hours the larger model answers slowly, times out or rejects requests. The second is censorship. On topics sensitive to the Chinese government the assistant starts writing an answer then deletes it before finishing, behaviour you can watch happen live and that goes away if you run the weights yourself. ## Pricing and plans The chat is entirely free. There is no Plus or Pro tier, the official site offers open access to the model, and the mobile app is listed as free in the stores. Daily fair-use quotas exist and can slow you down if you push hard, but no paid plan sits behind them. If you want to wire the model into your own software, the API does charge, and that is the other argument in DeepSeek's favour. Official rates start from $0.14 per million input tokens on V4-Flash and from $0.435 on V4-Pro, with an automatic discount when a request hits the cache. That is an order of magnitude below comparable models from Western providers, and it is worth checking the current rate before budgeting because the company has cut it more than once since launch. ## Where your data goes Its privacy policy states in plain words that it collects, processes and stores personal data in the People's Republic of China. That covers what you type, the files you upload and your chat history, and the data is used to train its models, though the same policy grants a right to opt out of that use. The consequences are already on record. Italy's data protection authority ordered an urgent limitation on processing Italian users' data in January 2025 over the lack of safeguards for transfers, and the app disappeared from the country's app stores. Australia, South Korea, Taiwan, the Czech Republic and several US agencies have banned its use on official devices. None of that makes DeepSeek unusable, but it does fence off where it belongs. For study, personal side projects or drafting, the risk is the one you take with any free service. For client data, medical records, legal documentation or anything covered by GDPR, the sensible answer is to skip the hosted version and look at the open weights on infrastructure you control, which is exactly the route the MIT license leaves open. ## Who it's for (and who it isn't) DeepSeek fits people who want serious reasoning without a monthly fee: students, freelancers who cannot justify three subscriptions at once, anyone wanting a second opinion alongside their main assistant, and technical teams cutting an API bill. It is also the obvious pick for running a large model on your own server. It does not fit people handling third-party data, nor anyone in public administration or a regulated sector, where data location gets decided before model quality does. It also does not fit anyone who needs memory, connectors, agents and images in one place, because [ChatGPT](https://serchai.com/en/reviews/chatgpt/) has a far more complete product wrapped around the model. ## Alternatives to DeepSeek If price was the draw, [Kimi](https://serchai.com/en/reviews/kimi/) and [Z.ai](https://serchai.com/en/reviews/glm/) play the same game of Chinese open-weight models at low rates, and they carry the same questions about where the data ends up. If you want a finished product from a company bound by European or US rules, [ChatGPT](https://serchai.com/en/reviews/chatgpt/) is the natural comparison. The full breakdown sits on our [DeepSeek alternatives](https://serchai.com/en/alternatives/deepseek/) page, and the category as a whole in our guide to the [best AI productivity tools](https://serchai.com/en/best-ai/productivity/). ## Frequently asked questions ### Is DeepSeek really free? Yes. The web chat and the mobile apps cost nothing and there is no paid plan for the assistant. Daily fair-use quotas can slow you down in long sessions, but no subscription sits behind them. What DeepSeek charges for is the developer API. ### Which model does DeepSeek run right now? The V4 family, released in April 2026 in two variants, V4-Pro and V4-Flash, with a one-million-token context window on the official services. The R2 model that forums kept predicting has not been published. ### Can I put work information into it? It depends on the information. Its policy states data is processed and stored in China and used to train its models, with an opt-out available. For client data, records or regulated documentation, the answer is no. ### Can I use it without sending anything to China? Yes, if you have the hardware or a provider you trust. The V4 weights are published under an MIT license, so you can download them and run them on your own infrastructure. That route drops both the data transfer and the official service's censorship filter. ## Alternatives - [ChatGPT](https://serchai.com/en/reviews/chatgpt/) — The AI assistant that covers the most ground. (4.1/5) - [Kimi](https://serchai.com/en/reviews/kimi/) — Moonshot's assistant with a million-token context window. (3.8/5) - [Z.ai (GLM)](https://serchai.com/en/reviews/glm/) — The assistant behind the GLM models, with its weights published openly. (4.1/5)