DeepSeek
: How to use it, features, and the business problems it solves
Add bookmark
What is DeepSeek?
DeepSeek is an open large language model (LLM) family, together with the AI chat and API services built on it, developed by DeepSeek (Hangzhou, China). Since October 17, 2026, the current model is DeepSeek-V4.1-Flash-1017, with weights released under the MIT License. Chat is available on the web and in the app, while the API supports a 1M-token context, image input, Thinking mode, and OpenAI / Anthropic compatibility, with pay-as-you-go rates halved outside peak hours. DeepSeek also offers the larger V4-Pro and the Search / Fetch APIs for web search.
Business problems it solves
About "DeepSeek"
What is DeepSeek
DeepSeek is an open large language model (LLM) family, together with the AI chat and API services built on top of it, developed by DeepSeek (Hangzhou, China). You can use its latest model through its chat products (web and mobile app), while developers get high performance at a low price through its open API platform. By releasing the weights of its main models openly, DeepSeek has drawn attention worldwide.
On October 17, 2026, the current model was updated to DeepSeek-V4.1-Flash-1017. The web version, the app, and the API all switched over the same day, and the weights are published on Hugging Face under the MIT License. For details, see our article What is DeepSeek-V4.1-Flash-1017 (in Japanese).
Source: DeepSeek official site (screen as of August 2026)
Current models
As of October 2026, the API offers two models: deepseek-flash and deepseek-v4-pro. Both handle long contexts of up to 1M tokens and can switch between Thinking mode (the default) and non-thinking mode.
01DeepSeek-V4.1-Flash-1017 (deepseek-flash)
The current mainstay. Released on October 17 as a minor update to V4.1-Flash (released September 10, 2026). It is a multimodal MoE model that accepts image input; the official model card describes it as having 552B backbone parameters, with 8B (prefill) / 16B (decode) parameters activated per token. On the code-repair benchmark DeepSWE v1.1 it scores 77.0 (self-reported by DeepSeek), up from 74.2 for the previous version.
02DeepSeek-V4-Pro-0813 (deepseek-v4-pro)
The large fourth-generation model (1.6T backbone parameters) that reached general availability on August 13, 2026. When V4.1-Flash was released, DeepSeek officially announced that it would keep offering V4-Pro on the API after September 14. It does not support image input (Vision).
03V3.2 / V3.1 / R1 / V3 and more
Earlier-generation models, such as the general-purpose V3 series and the reasoning-focused R1, remain published as research releases. The legacy API names deepseek-v4-flash and deepseek-v4-flash-vision-exp are still accepted, but those models have been retired and requests are served by V4.1-Flash-1017.
How to use
There are two ways to use DeepSeek: chat, and the developer API. You can also run the weights on your own servers.
Try it in chat
Sign up for the web version (chat.deepseek.com) or the app and start chatting. Since October 17 it runs on V4.1-Flash-1017.
Get an API key
Top up your balance on the open API platform and issue a key.
Call it by model name
Set model to deepseek-flash. You can call it in either OpenAI or Anthropic format.
Because the API is OpenAI-compatible, migrating is easy — you can often just swap the endpoint in your existing OpenAI-targeted code. The official docs also include setup guides for using it from Claude Code, Codex, OpenCode, and other tools.
Key features
01Thinking mode and Agent capabilities
Thinking mode is built in, so the model works through complex problems step by step. DeepSeek says V4.1-Flash-1017 improves on the previous version in agentic benchmarks such as code repair and terminal tasks, while also shortening task run time and output length.
021M-token context and image input
Load large volumes of documents or code at once, and generate up to 384K tokens of output. deepseek-flash also accepts image input (Vision).
03Search / Fetch API
On October 17, 2026, DeepSeek opened web search (Search) and page fetch (Fetch) APIs powered by its in-house search engine. Fetched pages are returned converted to Markdown.
04Developer-friendly API
Supports both OpenAI and Anthropic formats, with Tool Calls, JSON output, the Responses API, and Context Caching. On a cache hit, input pricing drops substantially.
Pricing
No paid plan is listed for the web or app chat; pricing applies to the API. The API is pay-as-you-go based on token usage, and rates differ by a factor of two depending on the time of day (as of October 18, 2026, in US dollars).
| Model | Time slot | Input (cache hit) | Input (cache miss) | Output |
|---|---|---|---|---|
| deepseek-flash | Off-peak | $0.003 | $0.15 | $0.6 |
| deepseek-flash | Peak | $0.006 | $0.3 | $1.2 |
| deepseek-v4-pro | Off-peak | $0.022 | $0.66 | $1.98 |
| deepseek-v4-pro | Peak | $0.044 | $1.32 | $3.96 |
Per 1M tokens. Source: DeepSeek API pricing page.
Peak hours are 01:00–04:00 and 06:00–10:00 UTC on weekdays (10:00–13:00 and 15:00–19:00 Japan time). All other hours, plus weekends and Chinese public holidays in full, are off-peak (half the peak rate).
| API | Price |
|---|---|
| Search (web search) | $0.75 per 1,000 successful requests |
| Fetch (page fetch) | $0.75 per 1,000 successful requests |
Search / Fetch calls made through DeepSeek Harness, Claude Code, or Codex are currently free. When you use the DeepSeek API from these tools, search and fetch are called on the server side; billing for those calls starts at 00:00 Beijing time on November 1, 2026. Charges are deducted from your prepaid balance (topped-up or granted). Always check the API pricing page for the latest rates.
How it differs from other LLMs
DeepSeek's strength is that it offers high-performance models openly and at a low price. It provides a 1M-token long context and Thinking mode, yet its API rates are far cheaper than those of the major proprietary models from overseas.
| Aspect | DeepSeek | Major overseas LLMs |
|---|---|---|
| Provider | DeepSeek (China) | OpenAI / Anthropic / Google, etc. |
| Price | Low (pay-as-you-go, off-peak and cache discounts) | Relatively higher |
| Openness | Weights released under the MIT License and others | Mostly closed |
| Compatibility | Supports OpenAI / Anthropic formats | Each vendor's own format, plus partial compatibility |
As general-purpose chat, its competitors are ChatGPT, Claude, and Gemini, while Qwen and Kimi, similarly open-minded models from China, occupy a close position. For developers who want to use high-performance models via API while keeping costs down, DeepSeek is a strong option.
Frequently asked questions (FAQ)
Q. Can I use DeepSeek for free?
No paid plan is listed for the web or app chat. Charges apply to API usage, which is pay-as-you-go based on tokens.
Q. What is the latest model?
DeepSeek-V4.1-Flash-1017, released on October 17, 2026. The web version, the app, and the API (deepseek-flash) all switched to it on the same day. The larger V4-Pro (deepseek-v4-pro) remains available on the API.
Q. Can I use it commercially?
The V4.1-Flash-1017 repository and weights are under the MIT License, which permits commercial use, modification, and redistribution (provided the copyright notice and license text are included). For chat and API outputs, the Terms of Use state that rights to outputs are transferred to the user.
Q. How is my data handled?
The provider is a Chinese company (DeepSeek). Its privacy policy states that collected information is stored on servers in the People's Republic of China, and its Terms of Use are governed by the laws of mainland China. The policy also gives users the right to refuse the use of their personal data for model training. If you handle sensitive information, you can also run the published weights in your own environment, so inputs are not sent outside.
Q. Can I migrate from existing OpenAI-targeted code?
Yes. The API supports the OpenAI format, so it's designed to make migration easy — you swap the endpoint and the API key.
The information on this page is based on what was published on the official site, the official API documentation, and the Hugging Face model card as of October 18, 2026. Models and pricing are subject to change, so please check the DeepSeek official site for the latest.
