DeepSeek is an AI assistant developed by Hangzhou DeepSeek Artificial Intelligence Co., Ltd.: you chat in natural language on the web (chat.deepseek.com) or in the mobile app to handle Q&A, writing, translation, coding and math reasoning. Its verifiable distinctions from most rivals: the chat product is free once you register; the underlying model weights are published on Hugging Face under permissive licenses such as MIT; and the API is priced well below leading overseas models while speaking both the OpenAI and Anthropic API formats. All dynamic facts in this article (model versions, prices, star counts) were checked against official pages on 2026-09-16.
At a Glance
- URL: https://chat.deepseek.com (assistant); https://www.deepseek.com (company and research)
- Type: conversational AI assistant (web + mobile app) plus a developer API platform
- Cost: web/app chat is free; the API is prepaid per token, with input starting at $0.15 per 1M tokens (official pricing, as of 2026-09-16)
- Sign-up: chatting requires an account; the sign-in page offers phone-number codes, password, Apple and WeChat QR login (tested), and the Terms of Use state registration by email or third-party account
- Interface language: the website and chat UI switch between Chinese and English (tested)
Background
According to TechCrunch's company profile, DeepSeek is backed by High-Flyer, a Chinese quantitative hedge fund, and started in 2023 as a dedicated AI lab under founder Liang Wenfeng; its first models (DeepSeek Coder, DeepSeek LLM, DeepSeek Chat) arrived in November 2023. The operating entity is Hangzhou DeepSeek Artificial Intelligence Co., Ltd. (Terms of Use, updated 2026-03-27), and the homepage footer shows the Zhejiang ICP registration 浙ICP备2023025841号.
Two releases put DeepSeek on the global map: DeepSeek-V3 in December 2024 and the R1 reasoning model in January 2025 — the latter benchmarked by the company against OpenAI's o1, while the DeepSeek app topped both the Apple App Store and Google Play free charts (TechCrunch). The official research news feed documents the lineage since: V3.1 (2025-08), V3.1-Terminus (2025-09), V3.2-Exp with DeepSeek Sparse Attention (2025-09-29, cutting API prices by over 50%), V3.2 (2025-12-01, with gold-medal results in IMO, CMO, ICPC and IOI 2025 per the announcement), V4 Preview (2026-04-24, opening the 1M-context era), and DeepSeek-V4.1-Flash (2026-09-10) — a 552B-parameter MoE model activating just 8B parameters for input and 16B for output, with native visual understanding and benchmark results the company says surpass its own V4-Pro flagship.
Chatting on the Web and App
The DeepSeek homepage doubles as the chat entry point (unauthenticated visitors are redirected to sign-in, as tested). The composer exposes two core, verifiable switches:
- DeepThink: the model produces a long reasoning chain before answering, suited to math, code and complex analysis;
- Search: the model retrieves web pages before answering, for time-sensitive questions. The terms caution that search improves accuracy "to some extent" and that inaccurate AI output remains an industry-wide problem that cannot be fully avoided (Terms of Use, §5.2).
Per the terms, models accept text, image and file inputs and produce text, tables and code; conversations can be shared through a unique link (the terms warn that publicly shared dialogues risk being scraped). The V4 announcement also mentions Expert Mode and Instant Mode on chat.deepseek.com (V4 Preview). The product line includes iOS and Android apps via the homepage's "Get App" entry.
API Platform and Pricing
The developer platform (platform.deepseek.com) bills per token against a prepaid balance, spending granted credit first. From the official pricing page as of 2026-09-16:
| Model | Version | Input per 1M tokens (cache miss) | Input (cache hit) | Output per 1M tokens |
|---|---|---|---|---|
| deepseek-flash | DeepSeek-V4.1-Flash | 0.15 off-peak | 0.003 | 0.60 |
| deepseek-v4-pro | DeepSeek-V4-Pro-0813 | 0.66 | 0.022 | 1.98 |
Peak hours are 01:00–04:00 and 06:00–10:00 UTC, Monday through Friday; all other hours are half price. API highlights: a 1M-token context window with up to 384K output tokens; switchable thinking and non-thinking modes; both OpenAI (https://api.deepseek.com) and Anthropic (https://api.deepseek.com/anthropic) endpoint formats; JSON output, tool calls, chat prefix completion and FIM completion; vision is supported on deepseek-flash only. A footnote states that, in response to user demand, V4-Pro remains available after September 14, 2026 with unchanged billing.
Open Models and the Developer Ecosystem
DeepSeek open-sources its flagship weights and a set of infrastructure projects under the deepseek-ai GitHub organization, mostly under MIT. As of 2026-09-16 (GitHub API data):
- deepseek-harness: about 225.7k stars — an open-source agent harness released in August 2026 (
npx @deepseek-ai/dsh webstarts its Web UI), built on an "everything-is-a-plugin" architecture and currently in developer preview; - DeepSeek-V3: about 104.5k stars; DeepSeek-R1: about 92.0k stars, MIT-licensed;
- Infrastructure and specialized models: FlashMLA, DeepEP, DeepGEMM, the 3FS distributed file system and DeepSeek-OCR, ranging from a few thousand to over 20k stars.
The V4.1-Flash weights released in September 2026 are also on Hugging Face under an MIT license (verified 2026-09-16), free for commercial use and derivative work. The team also maintains awesome-deepseek-integration (CC0), a catalog of DeepSeek API integrations with popular software.
Terms and Content Policy
Key points from the Terms of Use (2026-03-27):
- Ownership of output: you keep rights to your inputs, and DeepSeek assigns you its rights to the outputs. Outputs may be used for personal purposes, academic research, derivative products and even training other models (such as distillation), provided usage is lawful and complies with the terms. Outputs are not guaranteed to be unique — other users may receive similar results.
- AI labeling duty: when publishing generated content you must verify its accuracy and clearly mark it as AI-generated; removing or forging the platform's AI-content labels is prohibited.
- Training opt-out: subject to de-identification, inputs and outputs may be used minimally to improve the service; you can opt out by turning off "Improve the model for everyone" in settings.
- Use restrictions: no hateful, harassing, pornographic or violent content, nothing exploiting minors or instructing crime; no reverse engineering or scraping the service; accounts are non-transferable; minors under 18 need a guardian's consent.
- Jurisdiction: the terms are governed by the laws of mainland China, with disputes under the court where Hangzhou DeepSeek is registered; the terms also state the service is not guaranteed to be available in every jurisdiction and features may vary by region.
Where It Fits
- Everyday Q&A, writing, translation, coding help and math reasoning — especially problems that benefit from DeepThink's long reasoning chains.
- Academic and research work: the terms explicitly allow outputs to be used in research, and the 1M-token context suits long-paper Q&A and literature-review drafts.
- Developers integrating LLMs at low cost: dual OpenAI/Anthropic formats, cheap cache hits and time-of-day pricing favor high-frequency agent workloads.
- Teams needing self-hosting or further training: download the MIT-licensed weights and deploy them yourself.
Limitations
- Account required: there is no guest tier — unauthenticated visitors are sent straight to sign-in (tested). Login leans on phone numbers, WeChat and Apple, which can be inconvenient outside mainland China.
- Content boundaries: as a Chinese-developed service, the models are subject to China's internet regulations. TechCrunch reports the official app declines certain politically sensitive questions, and several governments and institutions (including U.S. Commerce Department bureaus, New York State and South Korea) have banned DeepSeek on government devices.
- Output reliability: the terms state outputs may contain errors or omissions, are for reference only, and do not constitute medical, legal or financial advice; human review is required for high-stakes use.
- Price changes: DeepSeek reserves the right to adjust pricing, so the figures above reflect only the pricing page on 2026-09-16; the API is prepaid, so estimate your usage before topping up.
- Regional variation: features may differ by jurisdiction, and some capabilities may be unavailable where you are.
Alternatives
- ChatGPT (listed on this site as chatgpt): OpenAI's conversational assistant with a broader plugin ecosystem and guest access for basic chats, but closed model weights.
- Perplexity (listed as perplexity): built around AI search with a stronger citation system, whereas DeepSeek's Search is one toggle inside a general chat product.
