DeepSeek’s release record is a model record, not a ranking record, but it now matters to search people for two reasons. Its open-weight models sit underneath a large share of Chinese AI answers, including Baidu, Tencent and many smaller apps, and since 17 October 2026 DeepSeek also runs its own web search and fetch APIs that decide which pages its models can see. This page lists DeepSeek model, API, pricing and search releases, newest first, with dates taken from DeepSeek’s own change log and news posts, model cards, papers, or Reuters and Bloomberg reporting. Where only the month is reliable, the row says so. Funding and outage news is in the essay below the table unless it changed what DeepSeek ships.
| Date | Update | What changed | Why it matters |
|---|---|---|---|
| 31 October 2026 | Search and Fetch billing for agent tools | From 16:00 UTC (midnight Beijing, 1 November), server-side search and fetch calls made through DeepSeek Harness, Claude Code and Codex are billed at $0.75 per 1,000 successful requests each. | Agent traffic that reads your pages through DeepSeek now carries a per-call cost, so retrieval is budgeted rather than free. |
| 20 October 2026 | DeepInfer open-sourced | DeepSeek’s own inference framework released under Apache 2.0, with deterministic sampling and Ascend support. | Cheaper self-hosting means more third-party answer engines running the same DeepSeek weights. |
| 17 October 2026 | Search API and Fetch API | In-house web index sold at $0.75 per 1,000 successful calls each. Results carry a rerank score and an authority tier, and the change log says the engine filters “SEO spam, GEO manipulation, and low-quality AIGC”. | The first Chinese AI lab to sell its own index. Visibility in DeepSeek answers now depends on DeepSeek’s crawler and authority scoring, not only on Bing or Baidu. |
| 17 October 2026 | DeepSeek-V4.1-Flash-1017 | Updated Flash checkpoint behind the deepseek-flash model ID, with open weights on Hugging Face. | The default API model, so it is the model most likely to be paraphrasing your content. |
| 29 September 2026 | DeepSeek Harness v0.2 | Open-source agent harness gains desktop apps; a 0.2.1 alpha followed on 3 October. | Harness is the client that later started paying for Search and Fetch calls, so it is a direct route from agents to the open web. |
| 10 September 2026 | DeepSeek-V4.1-Flash | 552B mixture-of-experts with a causal encoder–decoder (about 8B active parameters on input, 16B on output), native multimodal input, model ID deepseek-flash, lower prices from 04:00 UTC. The launch post said V4-Pro traffic would be routed to Flash from 14 September; the change log now says V4-Pro continues “in response to user demand”. | A smaller, cheaper model became the flagship in practice. The routing reversal is a reminder that model IDs can change underneath integrations. |
| 21 August 2026 | DeepSeek-V4-Flash-Vision-Exp | Experimental image-input version of V4-Flash on the API. | First vision input on the V4 line; screenshots and product images become readable inputs. |
| 13 August 2026 | DeepSeek-V4-Pro general availability | V4-Pro-0813 on the API with low, high and max thinking effort and the Responses API. Peak and off-peak pricing, with off-peak at half the peak rate, began at 16:00 UTC on 16 August. | Time-of-day pricing pushes bulk agent jobs, including crawling and summarising, into Chinese night hours. |
| 31 July 2026 | DeepSeek-V4-Flash-0731 | V4-Flash made generally available on the API, with Responses API support in public beta. | The cheap tier most developers build on. |
| 24 July 2026 | Legacy model names retired | deepseek-chat and deepseek-reasoner stopped working, as announced with V4. | Old integrations that quietly called these names broke or had to be migrated. |
| 24 April 2026 | DeepSeek-V4 Preview | V4-Pro (1.6T total, 49B active) and V4-Flash (284B total, 13B active), open weights, with a 1M-token context as the default on every service. | A million tokens means whole sites, catalogues or document sets can be read in one pass. |
| January 2026 (month only for OCR 2) | Engram and DeepSeek-OCR 2 papers | Engram conditional-memory paper on 12 January; DeepSeek-OCR 2 late in the month. | Research signals rather than product changes: better memory of facts and better reading of documents as images. |
| 1 December 2025 | DeepSeek-V3.2 and V3.2-Speciale | V3.2 adds thinking inside tool use; Speciale was a temporary API-only reasoning model, available until 15 December. | Reasoning while searching, rather than before or after it, is how agent search works now. |
| 20 October 2025 | DeepSeek-OCR | Open-source model that compresses text into vision tokens and decodes it back. | PDFs and scanned pages become cheap to ingest, which widens what can be cited. |
| 29 September 2025 | DeepSeek-V3.2-Exp | DeepSeek Sparse Attention for long context, and API prices cut by more than half. | Long-context reading got much cheaper, which favours retrieval-heavy products. |
| 22 September 2025 | DeepSeek-V3.1-Terminus | Fixes for mixed Chinese and English output and stronger Code Agent and Search Agent behaviour. | The search agent is the part that decides which pages get fetched. |
| 5 September 2025 | New API pricing | New price list from 16:00 UTC; the night-time discounts ended. | Ends the first off-peak scheme; a new one returned in August 2026. |
| 21 August 2025 | DeepSeek-V3.1 | One hybrid model with thinking and non-thinking modes, 128K context, Anthropic API format support and a stronger search agent. | DeepSeek could now be dropped into Claude-style tooling, which widened its reach in agents. |
| 28 May 2025 | DeepSeek-R1-0528 | Better benchmarks and front-end code, fewer hallucinations, JSON output and function calling. | Fewer invented facts and proper tool calls make R1 usable for search-backed answers. |
| 24 March 2025 | DeepSeek-V3-0324 | Better reasoning, front-end code and Chinese writing, and improved Chinese search-result analysis. | Directly about how DeepSeek reads and summarises Chinese search results. |
| 26 February 2025 | Off-peak discounts | 16:30–00:30 UTC daily: V3 50% off and R1 75% off. | The first time-of-day pricing, a template for later schemes. |
| 24–28 February 2025 | Open Source Week | A run of infrastructure repositories over five days, starting with FlashMLA and ending with the 3FS file system. | The serving stack became public, so other Chinese firms could run DeepSeek cheaply. |
| 27 January 2025 | App tops the US App Store | The DeepSeek app reached number one in the US; Nvidia lost about $589bn in market value that day. | The point where DeepSeek became a consumer answer engine, not only a developer API. |
| 20 January 2025 | DeepSeek-R1 | Open reasoning model under the MIT licence, with distilled smaller models. API at $0.55 per million input tokens ($0.14 on cache hits) and $2.19 output. | Within a month Baidu, Tencent and others had put R1 into their search and chat products. |
| 15 January 2025 | DeepSeek app | iOS and Android app on V3, free, with web search and DeepThink mode. | Web search in a free consumer app made DeepSeek a source of AI referrals. |
| 26 December 2024 | DeepSeek-V3 | 671B total parameters, 37B active, trained on 14.8T tokens, open weights. New prices from 8 February 2025. | The base model behind R1 and most later DeepSeek products. |
| 10 December 2024 | DeepSeek-V2.5-1210 and Internet Search | Last V2.5 checkpoint; Internet Search turned on at chat.deepseek.com, not on the API. | The start of DeepSeek answering with live web results. |
| 20 November 2024 | DeepSeek-R1-Lite-Preview | Reasoning preview in the chat interface. | An early look at the model family that went global two months later. |
| 5 September 2024 | DeepSeek-V2.5 | Chat and Coder models merged into one. | One general model for both writing and code. |
| 2 August 2024 | Context caching on disk | Repeated prompt prefixes are cached automatically; cache hits charged at $0.014 per million tokens, cutting costs by up to 90%. | Cheap repeat reads make large, re-used system prompts and documents affordable for agents. |
| 17 June 2024 | DeepSeek-Coder-V2 | Open code model with 338 programming languages and 128K context; the API switched to Coder-V2-0614 on 14 June. | Mainly a developer story, but it set DeepSeek’s reputation outside China. |
| 6 May 2024 | DeepSeek-V2 | 236B mixture-of-experts (21B active) with Multi-head Latent Attention, priced at about ¥1 per million input tokens. | The price that started China’s 2024 model price war; Baidu and Alibaba cut prices within weeks. |
| 29 November 2023 | DeepSeek LLM | 7B and 67B general models, open weights. | DeepSeek’s first general-purpose models. |
| 2 November 2023 | DeepSeek Coder | Open code models from 1.3B to 33B parameters. | DeepSeek’s first public release. |
Key model events
The first year is a lab finding its feet. DeepSeek Coder on 2 November 2023 and DeepSeek LLM on 29 November 2023 were respectable open models from a company spun out of the quant fund High-Flyer. The research papers that followed in early 2024, DeepSeekMoE, DeepSeekMath and DeepSeek-VL, are better treated as background than as rows, because the dates people quote are paper dates rather than releases. The turn comes on 6 May 2024 with DeepSeek-V2. A 236B mixture-of-experts model that only woke 21B parameters per token, plus a new attention design, let DeepSeek price an API at roughly a yuan per million input tokens. Within weeks Baidu made ERNIE Speed and Lite free and Alibaba cut Qwen prices. For search, the lasting effect was that large-model answers became cheap enough for every Chinese app to bolt on. The ERNIE model history shows the same price war from Baidu’s side.
The second half of 2024 is plumbing. Context caching on disk on 2 August cut repeat reads by up to 90%, V2.5 merged the chat and code models on 5 September, and on 10 December the V2.5-1210 post switched on Internet Search in the web chat. That last line is easy to miss, and it is the start of DeepSeek as a search product. V3 on 26 December gave the company a 671B base model, and R1 on 20 January 2025 put an open, MIT-licensed reasoning model next to it. The consumer app, released on 15 January with web search built in, reached the top of the US App Store a week after R1, and Nvidia lost about $589bn in a day. Chinese platforms did not wait: by mid-February Baidu Search, WeChat search and Tencent Yuanbao were all offering R1 answers. ReadIsi’s DeepSeek category follows the details.
2025 then settles into a steady cadence of checkpoints. V3-0324 is the one SEO people should remember, because DeepSeek’s own notes list better analysis of Chinese search results, and Artificial Analysis ranked it the top non-reasoning model at the time (our write-up). R1-0528 reduced hallucinations and added function calling. V3.1 on 21 August merged thinking and non-thinking modes and accepted the Anthropic API format, so DeepSeek could be dropped into tools built for Claude. V3.1-Terminus, V3.2-Exp and V3.2 kept improving the search agent and pushed long-context prices down. The pattern is consistent: each release makes it cheaper for an agent to fetch more pages and reason over them.
V4 on 24 April 2026 is the scale jump. V4-Pro has 1.6T parameters and V4-Flash 284B, and every DeepSeek service moved to a one-million-token context. The old deepseek-chat and deepseek-reasoner names were switched off on 24 July, V4-Flash and V4-Pro reached general availability on 31 July and 13 August, and from 16 August DeepSeek charged half price outside Chinese working hours. The most surprising release came on 10 September. V4.1-Flash beat the larger flagship on DeepSeek’s own tests, and the company said it would route V4-Pro calls to Flash. It later reversed that in the change log, so V4-Pro is still available; treat any coverage that says Pro was retired as out of date. Our speed comparison with Qwen is a useful reality check on the claims.
For search marketers the key date is 17 October 2026. DeepSeek launched Search and Fetch APIs backed by its own index, with a rerank score and an authority tier on every result, and the change log says the engine filters SEO spam, GEO manipulation and low-quality AI-generated content. That makes DeepSeek a search engine with its own quality rules, not just a model that reads someone else’s results. Billing for calls made through Harness, Claude Code and Codex began at 16:00 UTC on 31 October. Baidu answered on 30 October with a ¥4.9 per 1,000 calls promotion, and Tencent has opened the Sogou index the same way, so Chinese agent search is now a price war of its own.
The infrastructure releases explain how DeepSeek keeps prices low. It open-sourced Ascend tools with Huawei on 30 September and released the DeepInfer inference framework on 20 October. Its Harness desktop app and the DSec sandbox paper show where it is heading: agents that browse. Reliability is the weak point. A five-hour outage on 15 October ended with users credited for a week of spend.
Money and politics sit around all of this. Reuters reported DeepSeek’s first outside funding in June 2026. Bloomberg and Reuters reported on 23 October that a second round of about $15bn had closed with Tencent and CATL among the backers (our summary), and later reporting put it at ¥105.5bn from nearly 30 investors, including Meituan and Xiaohongshu (details). DeepSeek has not published those figures itself. In the US, the Department of Homeland Security has ordered federal agencies to remove DeepSeek, including from third-party products (our note). For an international site, the practical takeaways are simple: let DeepSeek’s crawler reach your Chinese pages, make authorship and sources obvious so the authority tier has something to score, and check DeepSeek answers for your brand after each major release. Wider coverage is in China AI and AI News.
Parallel histories: ERNIE and Baidu algorithm updates. Last reviewed 4 November 2026.