DeepSeek model timeline: V2 to V4.1, licenses and API names (2024–2026)
Live timeline · last updated · 1 entry in the update log
DeepSeek’s public record is an API changelog plus a set of model READMEs, and the two do not always use the same license words. This page follows the generation changes from V2 to V4.1, quotes the January 2025 sentence that names both R1 and V3, and points at the model card when a number lives there instead of here.
As of October 21, 2026 the newest dated DeepSeek build in the changelog is DeepSeek-V4.1-Flash-1017, on October 17, 2026, still called deepseek-flash. Original V3 weights were not MIT. R1 and the later V3-0324 news page are where MIT is stated.
Read on 2026-10-21: the DeepSeek API changelog, the DeepSeek-V2, DeepSeek-V3 and DeepSeek-R1 READMEs, the DeepSeek-V3.2 Hugging Face card, and the March 25, 2025 news page that states the MIT license. Benchmark deltas are DeepSeek’s. We did not run the models. An older news URL for the first V3 post redirected away, so we do not cite it.
Timeline
-
The DeepSeek-V2 README says the team released DeepSeek-V2 on this day. It gives 236 billion total parameters and 21 billion activated. Against DeepSeek 67B, the README says the model saves 42.5 percent of training costs, cuts the KV cache by 93.3 percent, and raises maximum generation throughput by 5.76 times. Those three comparisons are DeepSeek’s. The code license is MIT. The model license is a Model Agreement, not MIT. The same README dates DeepSeek-V2-Lite to May 16, 2024. We do not state a context length, because the sentences we are using do not give one.
-
The API changelog says DeepSeek V2 Chat and DeepSeek Coder V2 were merged and upgraded into DeepSeek V2.5. For backward compatibility, API users can call the new model as deepseek-coder or deepseek-chat. The changelog says the new model is ahead of the previous versions on general tasks and on code, and it prints DeepSeek-reported win rates: ArenaHard from 68.3 percent to 76.3 percent, AlpacaEval 2.0 length-controlled from 46.61 percent to 50.52 percent. Smaller API bumps in May, June and July 2024 stay in the gaps section. They are real changelog rows. They are not new generations.
-
The API changelog says deepseek-chat was upgraded to DeepSeek-V3. The API is unchanged. Callers pass model deepseek-chat. This row is the availability change. Parameter counts, context length and the weight license are on the DeepSeek-V3 README, quoted in the license section, not in this changelog paragraph.
-
The API changelog introduces deepseek-reasoner and says it is the new model DeepSeek-R1. The next sentence says you can invoke DeepSeek-V3 by specifying model deepseek-reasoner. Both sentences are on the page. We are not going to delete one of them to make the paragraph consistent. The README’s size table and MIT badge are a different page. They are quoted in the license section as the R1-Zero row, not as a clarification of this contradictory pair.
-
The API changelog upgrades deepseek-chat to DeepSeek-V3-0324 and prints DeepSeek-reported deltas: MMLU-Pro 75.9 to 81.2, GPQA 59.1 to 68.4, AIME 39.6 to 59.4, and LiveCodeBench 39.2 to 49.2. It also describes better front-end code, Chinese writing, multi-turn rewriting, translation, Chinese search and function calling. The MIT claim is not in this paragraph. The license section cites the separate news page that does state MIT.
-
The API changelog upgrades deepseek-reasoner to DeepSeek-R1-0528. Pass@1 figures, DeepSeek-reported: AIME 2025 from 70.0 to 87.5, GPQA from 71.5 to 81.0, LiveCodeBench v6 from 63.5 to 73.3, and Aider from 57.0 to 71.6. The same note adds JSON output and function calling, with Tau-bench at 53.5 on Airline and 63.9 on Retail. It also says complex reasoning may consume more tokens than the legacy R1. The token warning is the part that changes a bill even when the model name looks like a small revision.
-
The API changelog upgrades both deepseek-chat and deepseek-reasoner to DeepSeek-V3.1. Chat is the non-thinking mode. Reasoner is the thinking mode. DeepSeek calls the design a hybrid reasoning architecture and says V3.1-Think answers in significantly less time than R1-0528, without a number on that comparison. DeepSeek-reported scores on the same row: SWE-bench Verified 66.0, SWE-bench Multilingual 54.5, Terminal-bench 31.3. A separate news page has been cited elsewhere for a 128K context and a continued-pretraining token count. Those figures are not in the changelog paragraph we re-read, so they are not in this node.
-
The API changelog upgrades both names to DeepSeek-V3.2, again with chat as non-thinking and reasoner as thinking. DeepSeek-V3.2-Speciale is a temporary endpoint, base URL https://api.deepseek.com/v3.2_speciale_expires_on_20251215, same price as V3.2, no tool calls, available until December 15, 2025, 15:59 UTC. The Hugging Face license name and the missing parameter count are in the license section. Gold-medal claims are not in this changelog paragraph, so they are not repeated here.
-
The API changelog says the API supports V4-Pro and V4-Flash, via the OpenAI Chat Completions interface and the Anthropic interface. The model names are deepseek-v4-pro and deepseek-v4-flash. The base URL stays the same. The same paragraph says the legacy names deepseek-chat and deepseek-reasoner will be discontinued in three months, on July 24, 2026, and that during the then-current period those names point at the non-thinking and thinking modes of deepseek-v4-flash. That is a schedule written on April 24. It is not, by itself, a later confirmation that the names went dark on July 24.
-
The API changelog releases DeepSeek-V4.1-Flash and calls it the smallest model in the new architecture family, with native multimodal visual understanding. The API name to call is deepseek-flash. The previous names deepseek-v4-flash and deepseek-v4-flash-vision-exp are temporarily routed to V4.1 Flash. DeepSeek says it will keep serving V4-Pro after September 14, 2026, with billing unchanged. DeepSeek-reported scores on the row include GPQA Diamond 90.9, HLE 36.8 (39.1 on the pure-text subset) and a Codeforces rating of 3471. Parameter counts and the price table live on the model card and the pricing page. Our October 17 post reads those cards. This node does not copy figures we did not re-fetch from them.
-
The API changelog upgrades V4.1 Flash to DeepSeek-V4.1-Flash-1017 on the web, the app and the API. The calling method and the pricing stay the same. The name to use is deepseek-flash. The same day’s changelog opens the in-house search API. What changed in the weights, and what the card says about size, is written up separately on this site from the model card. The changelog also prints a new benchmark list. Treat that list as DeepSeek-reported. We are not pasting it again next to a post that already argues over the card.
How we date a DeepSeek release
The API changelog at api-docs.deepseek.com is the date source for API model swaps. Each of those nodes links that page. The page is one long log, so the date in the node is the heading we are quoting, not a claim that the whole page was published that day.
Weight and license facts come from the repository README or the Hugging Face card named in the node. A Hugging Face createdAt timestamp is not a release date. For DeepSeek-V2 the card’s created time we saw was earlier than the README’s May 6, 2024 release line, which is why the README wins.
The last check is October 21, 2026. The October 17 changelog row is the newest model row we are using. Pricing pages move. If a price is not in the sentence we quote, it is not on this page.
The license is not one story
The DeepSeek-V2 README splits code and weights: code under MIT, the model under a Model Agreement. The DeepSeek-V3 README does the same, and it is the page that states 671 billion total parameters, 37 billion activated, multi-head latent attention, DeepSeekMoE and a context of up to 128K. Calling either checkpoint “the MIT model” is wrong even though the code repository is MIT.
R1 is the break. The DeepSeek-R1 README shows an MIT badge, and the table row we are willing to quote is DeepSeek-R1-Zero at 671 billion total, 37 billion activated, 128K context. On March 24, 2025 the changelog upgrades the chat model to V3-0324 and does not say MIT in that paragraph. The V3-0324 news page says the models are now released under the MIT License, just like DeepSeek-R1. That is the sentence to cite for V3-0324’s license, not the changelog deltas.
The DeepSeek-V3.2 card lists license mit, titles the model “DeepSeek-V3.2: Efficient Reasoning & Agentic AI,” and names the base model DeepSeek-V3.2-Exp-Base. The head we read does not list a parameter count. V4.1 sizes are on the model card covered by our V4.1 Flash 1017 timeline. Copying them here without opening the card again would freeze a number we did not re-check on this pass.
API names, and the sentence that points two ways
For a long stretch, two ids did all the work. deepseek-chat was the non-thinking model. deepseek-reasoner was the thinking model. V2.5, V3, V3-0324, R1, R1-0528, V3.1 and V3.2 all landed as upgrades behind those ids. An application that pinned an id rather than a snapshot silently changed model on each of those days.
January 20, 2025 is the row to read twice. The changelog says deepseek-reasoner is the new model DeepSeek-R1, and that you can invoke DeepSeek-V3 by specifying model deepseek-reasoner. A migration guide that picks only the first clause will tell readers the wrong model. We leave both clauses in the node.
April 24, 2026 replaces that pair with deepseek-v4-pro and deepseek-v4-flash, and it writes a discontinuation date of July 24, 2026 for the old names. September 10, 2026 introduces deepseek-flash for V4.1 Flash, routes the old V4 Flash names to it for compatibility, and says V4-Pro keeps being served after September 14. October 17 keeps the name deepseek-flash and says pricing is unchanged. Whether deepseek-chat still answered on July 24 is a later observation. The April 24 text is a plan.
Rows on the changelog that are not nodes
The log also records DeepSeek-V2-0517 on May 17, 2024, Coder-V2-0614 on June 14, V2-0628 on June 28, Coder-V2-0724 on July 24, and V2.5-1210 on December 10, 2024. It records V3.1-Terminus on September 22, 2025 and V3.2-Exp on September 29, 2025. In 2026 it records V4-Flash public beta on July 31, V4-Pro general availability on August 13, and V4-Flash-Vision-Exp on August 21 under the name deepseek-v4-flash-vision-exp. The August 13 row says prices change at 16:00 UTC on August 16, 2026, with off-peak at half the peak price, and that thinking effort can be low, high or max. The August 21 row reports Terminal Bench 2.1 at 83.9, DeepSeek’s figure.
Those rows are real. They are revisions, an experimental vision id, or a pricing clock, and the generation nodes above already name the model family they belong to. The July 31 note says V4-Flash-0731 keeps the preview’s architecture and size and was only re-post-trained. That is why it is not a second architecture node. Full benchmark tables on those rows are DeepSeek-reported and are left on the changelog.
Where the October 17 card is written up
Size, file layout and the price in force for deepseek-flash are the subject of the DeepSeek V4.1 Flash 1017 timeline. This page stops at the changelog: the build id, the unchanged API name, the unchanged price, and the search API opening on the same day. If the two pages ever disagree, the model card wins on weights and the changelog wins on the day the API name moved.
What we left out
- A parameter count for V3.2. The card head we read does not print one.
- V3.1’s context length and continued-pretraining token count. They were not in the changelog paragraph re-read for this page.
- A confirmation that deepseek-chat and deepseek-reasoner stopped on July 24, 2026. April 24 only schedules that day.
- Gold-medal claims for V3.2. They are not in the changelog paragraph quoted here.
- The first V3 news URL that redirected away from the announcement. The December 26 changelog row and the V3 README are the sources we kept.
Related timelines
The same format covers Claude, OpenAI, Gemini, Grok and Meta Muse. Labs covered in this batch are Moonshot Kimi, Mistral and Alibaba Qwen. The build-level writeup is the V4.1 Flash 1017 timeline. The Front blog index lists every post.
Keep reading on Front
Every post, newest first, is on the Front blog index.
- DeepSeek V4.1 Flash 1017 timeline: DeepSeek’s V4.1 Flash release, its Oct 17 update, prices and parameter counts.
- Kimi model timeline: Moonshot’s Kimi line from K1.5 to K3.1 Pro, with licenses and context.
- Qwen model timeline: Qwen from Qwen-7B to Qwen3-Coder, with licenses and context.
- Mistral model timeline: Mistral from 7B to Large 4, with the license on each release.
FAQ
What is the latest DeepSeek model?
As of October 21, 2026 the newest changelog row is DeepSeek-V4.1-Flash-1017, dated October 17, 2026. The API name on that row is deepseek-flash.
Is DeepSeek-V3 MIT licensed?
The original V3 README puts the code under MIT and the model under a Model Agreement. A DeepSeek news page says the V3-0324 models are released under the MIT License, like DeepSeek-R1. Those are different checkpoints.
Which API name should I call?
For the October 17, 2026 build the changelog says deepseek-flash. deepseek-v4-pro and deepseek-v4-flash are the April 24, 2026 names. deepseek-chat and deepseek-reasoner are the older pair.
Why does the January 20, 2025 note mention both R1 and V3?
Because the changelog does. It says deepseek-reasoner is DeepSeek-R1, and that you can invoke DeepSeek-V3 with model deepseek-reasoner. We quote both clauses.
Update log
- — First published. Checked the DeepSeek API changelog and the V2, V3, R1 and V3.2 model pages through October 21, 2026.