The Full DeepSeek History: Every Model and Milestone, 2023-2026

If you have only met DeepSeek through the January 2025 R1 headlines, the rest of the story is missing. The full DeepSeek history starts with a lab linked to a quantitative hedge fund, runs through a string of open-weight model releases and API price changes, and — as of 24 October 2026 — reaches two current API models from two different releases: deepseek-flash, served by DeepSeek-V4.1-Flash-1017 (17 October 2026), and deepseek-v4-pro, served by DeepSeek-V4-Pro-0813 (13 August 2026). This guide walks the timeline release by release, dated from DeepSeek’s API Change Log, news pages, GitHub releases and papers; market and regulatory events are attributed to the outlets that reported them.

Before DeepSeek: the High-Flyer connection

DeepSeek did not start as a consumer AI brand. According to The Guardian (27 January 2025), it was founded in 2023 by Liang Wenfeng, who runs the hedge fund High-Flyer, and is based in Hangzhou. The connection also shows up in DeepSeek’s own papers: the DeepSeek LLM, DeepSeek-Coder, DeepSeekMath and DeepSeek-V2 technical reports all say their models were trained with HAI-LLM, a framework they cite as High-Flyer’s (2023). DeepSeek does not publish a corporate history, so the record below starts with the first models in late 2023.

2023: the first models — Coder and LLM

The first releases were modest by 2026 standards but established a pattern that held for most of the timeline: publish the weights, then a technical report.

• DeepSeek Coder (November 2023) — code models in 1B, 5.7B, 6.7B and 33B sizes, trained from scratch on 2T tokens (87% code, 13% natural language in English and Chinese), according to the GitHub README (repository published 2 November 2023). The technical report followed on arXiv on 25 January 2024.
• DeepSeek LLM (November 2023) — general-purpose dense models in 7B and 67B sizes, base and chat, trained on 2 trillion tokens in English and Chinese; published on GitHub on 29 November 2023. The paper, DeepSeek LLM: Scaling Open-Source Language Models with Longtermism, was posted to arXiv on 5 January 2024.

Both lines published their code under MIT, while the model weights came under a separate DeepSeek Model License that permits commercial use. Later releases (R1 onwards) moved the weights to MIT as well.

2024: MoE, Math, V2 and context caching

2024 was the year DeepSeek’s research bets started compounding:

• DeepSeekMoE (January 2024) — a 16B Mixture-of-Experts model (base and chat); paper on arXiv on 11 January 2024.
• DeepSeekMath (February 2024) — 7B base, instruct and RL models; paper on arXiv on 5 February 2024.
• DeepSeek-V2 (6 May 2024) — a 236B-parameter MoE model with 21B active per token. The paper (arXiv, 7 May 2024) introduced Multi-head Latent Attention (MLA) alongside the DeepSeekMoE design and, by DeepSeek’s account, cut training costs by 42.5%, the KV cache by 93.3% and raised maximum generation throughput 5.76-fold compared with DeepSeek 67B.
• DeepSeek-Coder-V2 (June 2024) — the API’s deepseek-coder model was upgraded to DeepSeek-Coder-V2-0614 on 14 June 2024; the open models (16B and 236B MoE) extended language support from 86 to 338 programming languages and context from 16K to 128K.

The API side moved just as fast. The Change Log records JSON mode, function calling, Chat Prefix Completion (Beta) and FIM completion (Beta) on 25 July 2024; context caching on disk on 2 August 2024, which DeepSeek said reduced prices “by another order of magnitude”; and on 5 September 2024 the merger of the V2 chat and Coder V2 models into DeepSeek V2.5, served under both deepseek-chat and deepseek-coder. DeepSeek-V2.5-1210 followed on 10 December 2024.

DeepSeek-V3 (December 2024)

On 26 December 2024 the API’s deepseek-chat model was upgraded to DeepSeek-V3. DeepSeek’s announcement described a 671B-parameter MoE model with 37B active parameters, trained on 14.8T tokens. The technical report (arXiv, 27 December 2024) added an auxiliary-loss-free load-balancing strategy and a multi-token prediction training objective, and put the full training run at 2.788M H800 GPU hours — $5.576 million at an assumed rental price of $2 per GPU hour. The report notes that this excludes the costs of prior research and ablation experiments.

January 2025: R1 and the “Sputnik moment”

On 20 January 2025 DeepSeek released DeepSeek-R1 on its website and API (model='deepseek-reasoner'). The release note claimed performance on par with OpenAI-o1 on math, code and reasoning tasks (DeepSeek-reported), put the code and weights under the MIT License, and released six smaller distilled models. R1 has 671B total and 37B active parameters and was trained from DeepSeek-V3-Base; the technical paper went up on arXiv on 22 January 2025.

The market reaction came a week later. On 27 January 2025 The Guardian reported that the DeepSeek assistant had topped Apple’s App Store in the US and UK over the weekend, and quoted venture capitalist Marc Andreessen calling R1 AI’s “Sputnik moment”. The same day, CNBC reported that Nvidia shares fell 17%, a market-cap loss of close to $600 billion and the biggest one-day drop ever for a US company.

One detail to keep straight: the famous “under $6 million” figure belongs to V3, not R1. DeepSeek’s R1 paper in Nature (published 17 September 2025) lists R1’s own training cost in Supplementary Table 4 as 147K H800 GPU hours, or $294,000 at $2 per GPU hour — $202K for R1-Zero, $10K for SFT data creation and $82K for R1. Both numbers are DeepSeek’s; they belong to two different training runs.

For a closer look at the model itself, see our DeepSeek R1 review and the DeepSeek R1 vs OpenAI o1 comparison.

Mid-to-late 2025: regulatory blowback and a quieter year of iteration

R1’s success did not come without friction. On 30 January 2025 Italy’s data-protection authority, the Garante, ordered an urgent limitation of the processing of Italian users’ data by DeepSeek. The next day the Associated Press reported that Texas had banned DeepSeek from state-issued devices; Taiwan’s public sector (Radio Free Asia, 3 February 2025) and Australian government systems (Bloomberg, 4 February 2025) followed, and South Korea suspended new app downloads from 15 February 2025 (Reuters via NBC News, 17 February 2025). Most of these measures target government use rather than private users. In the US, a Department of Homeland Security order published in the Federal Register on 20 October 2026 now also excludes DeepSeek products from federal civilian agency IT. See our guides to DeepSeek availability by country and DeepSeek US restrictions for the current picture.

On the model side, the year was one of steady upgrades rather than a new reasoning flagship. The Change Log upgraded deepseek-chat to DeepSeek-V3-0324 on 24 March 2025 and deepseek-reasoner to DeepSeek-R1-0528 on 28 May 2025 (DeepSeek-reported AIME 2025 up from 70.0 to 87.5, plus JSON output and function calling). It lists no model called R2. Other shifts that mattered for developers:

• 21 August 2025 — DeepSeek-V3.1 shipped as one model with thinking and non-thinking modes (switched by the “DeepThink” button in the chat), 128K context for both API names, and support for the Anthropic API format.
• 5 September 2025 — per the V3.1 announcement, new API pricing started and the off-peak discounts ended at 16:00 UTC. A time-based discount returned in August 2026 (see below).
• 22 and 29 September 2025 — DeepSeek-V3.1-Terminus, then the experimental DeepSeek-V3.2-Exp.
• 1 December 2025 — DeepSeek-V3.2 became the model behind deepseek-chat and deepseek-reasoner, with V3.2-Speciale on a temporary API endpoint until 15 December 2025. V3.2 was DeepSeek’s first model to integrate thinking into tool use.

2026: from the V4 Preview to V4.1-Flash-1017

2026 brought six model releases in six months:

• 24 April 2026 — DeepSeek-V4 Preview. deepseek-v4-pro and deepseek-v4-flash arrived on the API with both the OpenAI ChatCompletions and Anthropic interfaces, and 1M context became the default across DeepSeek’s official services. The April preview models were V4-Pro (Preview) with 1.6T total / 49B active parameters and V4-Flash (Preview) with 284B total / 13B active — figures that belong to the preview only. The legacy names deepseek-chat and deepseek-reasoner were pointed to V4-Flash’s non-thinking and thinking modes and were scheduled for retirement on 24 July 2026 at 15:59 UTC.
• 31 July 2026 — DeepSeek-V4-Flash-0731. A public beta of the official (non-preview) V4-Flash, with the same architecture and size as the preview, only re-post-trained.
• 13 August 2026 — DeepSeek-V4-Pro GA (DeepSeek-V4-Pro-0813) on the app, web and API, with low / high / max thinking effort and native OpenAI Responses API support. DeepSeek also announced peak and off-peak pricing, with off-peak at half the peak rate, effective 16:00 UTC on 16 August 2026.
• 21 August 2026 — DeepSeek-V4-Flash-Vision-Exp, an experimental vision model on the API, together with a free Files API.
• 10 September 2026 — DeepSeek-V4.1-Flash as deepseek-flash, with native image understanding; DeepSeek’s news page describes it as a 552B-parameter MoE with 8B active parameters for input and 16B for output. V4-Flash and V4-Flash-Vision-Exp were retired (their names now route to V4.1-Flash) and API prices were cut from 04:00 UTC that day. The news page announced that deepseek-v4-pro would be routed to V4.1-Flash from 14 September, but the Change Log says DeepSeek decided to keep serving V4 Pro after that date with billing unchanged.
• 17 October 2026 — DeepSeek-V4.1-Flash-1017 on the API, web and app, with the calling method and pricing unchanged, and the Search and Fetch APIs opened to everyone at $0.75 per 1,000 calls each.

Today the API lists deepseek-flash (DeepSeek-V4.1-Flash-1017, the only current model with vision) and deepseek-v4-pro (DeepSeek-V4-Pro-0813), both with a 1M-token context, up to 384K output tokens and thinking mode on by default. For request details see the DeepSeek API documentation.

The complete release timeline at a glance

Pricing history — what each generation actually charged

Pricing is a useful way to read DeepSeek’s strategy, but the line has not only gone down: current Flash output costs more per token than V3.2 did. Historical rates below come from DeepSeek’s release notes; none of them is current.

Current rates as of October 2026, from the official pricing page: off-peak is exactly half of peak ($0.30 ÷ 2 = $0.15; $1.20 ÷ 2 = $0.60; $1.32 ÷ 2 = $0.66; $3.96 ÷ 2 = $1.98). Peak hours are 01:00–04:00 and 06:00–10:00 UTC, Monday to Friday, excluding Chinese public holidays; all other hours are off-peak. The milestones in between: cache-hit pricing arrived with context caching on 2 August 2024; the off-peak discounts ended on 5 September 2025; peak/off-peak pricing returned on 16 August 2026; and prices were cut again on 10 September 2026. For worked cost examples see our DeepSeek API pricing guide.

The architecture story underneath

The timeline reads as a series of efficiency gains because the research steps compound, and each is documented in a paper or release note:

• DeepSeekMoE (January 2024) and Multi-head Latent Attention (V2, May 2024) — the two designs the V3 report says were “thoroughly validated in DeepSeek-V2” and carried forward.
• Auxiliary-loss-free load balancing and multi-token prediction — introduced in the V3 report (December 2024).
• Large-scale reinforcement learning — the R1 paper’s central claim is that reasoning ability can be incentivised through pure RL, without human-labelled reasoning trajectories (R1-Zero).
• Hybrid thinking — V3.1 (August 2025) put thinking and non-thinking modes into one model; V3.2 (December 2025) added thinking inside tool use.
• Token-wise compression plus DeepSeek Sparse Attention (DSA) — the attention design DeepSeek credited for cheap 1M-token context in the April 2026 V4 Preview announcement.
• DSpark speculative decoding — the module the V4-Pro-0813 model card says is attached to the V4-Pro structure.
• Asymmetric Causal Encoder–Decoder — the September 2026 V4.1-Flash design, which DeepSeek says needs a quarter of the previous generation’s HBM for the KV cache.

Where to go next

If this timeline is the first DeepSeek piece you have read, three follow-ups are worth your time depending on your goal:

• To use the current generation — see our overview of DeepSeek V4, the page for DeepSeek V4-Pro and the history of the now-retired DeepSeek V4-Flash; for what DeepSeek has announced next, see the DeepSeek roadmap.
• To compare against Western incumbents — read DeepSeek vs ChatGPT and DeepSeek vs Claude.
• For the broader catalogue of model write-ups, the DeepSeek models hub lists releases individually.