DeepSeek launched its own web search engine and API endpoints in October 2026, transitioning from third-party providers to an in-house crawling and indexing stack.
DeepSeek's Search API, launched on 17 October 2026, runs on a web search engine that DeepSeek built itself. DeepSeek says the engine indexes hundreds of billions of web pages and handles hundreds of millions of search requests a day.
DeepSeek's own technical papers describe the company's crawler, and its 2026 job ads describe a full search stack: a crawler, an index, ranking and spam detection. The engine is new. In December 2025 DeepSeek still tested its search agents on a commercial search API, and in 2025 the DeepSeek app's web search ran on Bocha, a Chinese search API provider.
DeepSeek released two endpoints on api.deepseek.com, documented here. Search (POST /v1/tools/search) returns web results for a query. Fetch ( POST /v1/tools/fetch) downloads one web page and returns it as Markdown.
Each endpoint costs $0.75 per 1,000 successful requests. Empty results, failures and timeouts are not billed, and each endpoint is limited to 5 queries per second per account. Search and Fetch calls made from DeepSeek Harness, Claude Code and Codex are free until 00:00 Beijing time on 1 November 2026.
| Field | Direction | What it does |
|---|---|---|
query |
Request | The search text, 1 to 128 characters |
count |
Request | Number of results, 1 to 50, default 10 |
ori_question |
Request | The user's original question, used to improve query understanding |
localization |
Request | An IP address used to localise results |
need_content |
Request | Whether to return each page's full text |
rerank_score |
Response | A score from 0 to 3 combining relevance, authority and freshness; 1.5 or above counts as high quality |
authority |
Response | A source authority tier from 0 to 4 |
snippet ,summary ,content |
Response | A 100 to 200 character snippet, a 600 to 800 character summary, and the full page text |
title ,url ,site_name ,site_icon ,date |
Response | Page metadata |
Queries accept search operators familiar from Google: site:, -site:, -term, exact phrases in quotes, lang:, after: and before:.
The API documentation says Search is powered by DeepSeek's in-house search engine. DeepSeek's launch changelog adds that the same engine runs search in the DeepSeek app and in the API.
| Claim | DeepSeek's figure |
|---|---|
| Pages indexed | Hundreds of billions |
| Search requests handled | Hundreds of millions a day, worldwide |
| Languages | Multilingual |
| Filtering | Removes SEO spam, GEO manipulation and low-quality AI-generated content |
GEO (generative engine optimisation) means editing web content so that AI answer engines cite it. DeepSeek lists it next to SEO spam as something its index filters out.
DeepSeek's technical reports show search moving from a commercial API in late 2025 to DeepSeek's own crawler by September 2026.
| Date | Paper | What it says about search |
|---|---|---|
| Sep 2026 | DeepSeek-V4.1-Flash | Refers to "our crawling system". DeepSeek found it favoured text pages and rebuilt it from Common Crawl to cover more images and other media |
| Apr 2026 | DeepSeek-V4 | Search agents run in an in-house harness with web search and Python tools. The app uses search and fetch tools. The model has special tokens for writing search queries and for judging how authoritative a source a question needs |
| Dec 2025 | DeepSeek-V3.2 | Search agents were trained on real web search APIs and evaluated with "a standard commercial search API" |
| Feb 2024 | DeepSeekMath | Web training data came from Common Crawl: a classifier picked 35.5 million maths pages out of 40 billion HTML pages |
The V4 authority token matches the authority field the Search API returns. Common Crawl is a free public archive of the web; DeepSeek uses it as the starting list of pages for its own crawler.
DeepSeek's 2026 job ads describe every part of a web search engine, from crawling to ranking. Bloomberg reported in January 2026 that more than a dozen DeepSeek postings sought engineers to build a multilingual, multimodal AI search engine.
| Date | Role | What the ad describes |
|---|---|---|
| 18 Sep 2026 | Pre-training data engineer, collection pipeline | Link-quality prediction, crawl quotas per site, and fresh seed URLs so that search results stay current |
| 27 Jun 2026 | AI search algorithm / architecture engineer | The "DeepSeek search engine": query understanding, retrieval, ranking, index selection, and an offline index-building pipeline for hundreds of billions of data items and tens of millions of daily users |
| 27 Jun 2026 | Pre-training data engineer | A distributed, fault-tolerant crawler; removing low-quality site farms and duplicate mirror sites |
| 3 Mar 2026 | Deep learning R&D engineer | Query rewriting, term weighting, ranking, index quality and spam detection across hundreds of billions of pages and trillions of links |
| 3 Mar 2026 | Search algorithm engineer | Reranking on authority, freshness, page quality and click behaviour |
| 3 Mar 2026 | Full-stack engineer, data infrastructure | Building the web page library and the index library |
| 29 Jan 2026 | Web data collection engineer | A distributed crawler that follows robots.txt |
Three of these ads are on the careers portal of High-Flyer, DeepSeek's parent company. The other four are on BOSS直聘 (BOSS Zhipin), a Chinese job site. The text above was read from copies on the aggregator DataHub, each of which links to its original posting.
Two ads ask for experienced search engineers. One prefers candidates with algorithm work at a mainstream large-scale search engine, and another prefers experienced search experts or team leads.
In 2025 the DeepSeek app's web search ran on Bocha (博查), a Chinese company that sells a search API. Bocha's CTO told 每日经济新闻 (National Business Daily) in March 2025 that DeepSeek had connected to Bocha's API before DeepSeek's January 2025 surge in users.
Bocha's own figures from March 2025 were 30 million API calls a day and an index of close to 10 billion web pages. Its GitHub README still calls Bocha DeepSeek's official web-search supplier.
| Bocha (2025) | DeepSeek Search (2026) | |
|---|---|---|
| Pages indexed | Close to 10 billion | Hundreds of billions |
| Response format | Bing-style webPages.value list withname ,siteName ,datePublished |
Flat results list withrerank_score andauthority |
No DeepSeek or Bocha statement says when the arrangement ended, and no source links Bocha to the October 2026 API.
DeepSeek runs a web crawler that identifies itself in server logs as:
Mozilla/5.0 (compatible; DeepSeekBot/1.0; +https://www.deepseek.com/)
The bot directory Known Agents classes DeepSeekBot as a high-volume AI data scraper and reports that 11% of the top 1,000 websites block it. DeepSeek's training-data summary for the EU says its crawler respects robots.txt. Bot directories describe DeepSeekBot as a crawler for training data, and DeepSeek has not named a separate crawler for its search index.
DeepSeek's storage is large enough to hold a web index. Its Fire-Flyer cluster paper describes 180 storage nodes with 2,880 NVMe SSDs, more than 20 PiB of capacity and 8 TB/s of measured read throughput, running DeepSeek's own 3FS file system.
DeepSeek's open-source agent harness, deepseek-harness, registers DeepSeek's engine as a search provider named deepseek-official, next to providers for Exa and Perplexity.
DeepSeek's English privacy policy, last updated on 10 February 2026, says DeepSeek integrates "third-party APIs to provide search services". The Chinese version of the same date names no search vendor. Both versions predate the October launch by eight months, and neither says whether any search traffic still goes to an outside provider.
No DeepSeek employee has a public record of engineering work at Baidu, Bing, Google, Sogou or another web search engine. No search or information-retrieval paper lists a DeepSeek affiliation.
No one outside DeepSeek has published a measurement of the index size, a comparison of its results with other engines, or the user agent and IP addresses the Fetch endpoint uses.