DeepSeek release history.

DeepSeek has shipped 15 major model releases and 2 other major launches since DeepSeek arrived in November 2023. The latest model recorded here is DeepSeek-V4.1-Flash, announced 10 Sep 2026. Every entry is dated from DeepSeek’s own announcement and links to it.

17 MAJOR RELEASES2023–2026CHECKED 23 SEP 2026
NEWEST FLASH
DeepSeek-V4.1-Flash · 10 Sep 2026
NEWEST PRO
DeepSeek-V4-Pro general availability · 13 Aug 2026
NEWEST JANUS
Janus-Pro · 27 Jan 2025
§ 01 — LINEAGE

How the DeepSeek models evolved.

PRODeepSeek-V4 Preview — 2026-04-24V4 previewDeepSeek-V4-Pro general availability — 2026-08-13V4FLASHDeepSeek-V4 Preview — 2026-04-24V4 previewDeepSeek-V4-Flash official release — 2026-07-31V4DeepSeek-V4.1-Flash — 2026-09-10V4.1JANUSJanus-Pro — 2025-01-27ProLLM & V-SERIESDeepSeek LLM 7B and 67B — 2023-11-29LLM 67BDeepSeek-V2 — 2024-05-06V2DeepSeek-V2.5 — 2024-09-05V2.5DeepSeek-V3 — 2024-12-26V3DeepSeek-V3.1 — 2025-08-21V3.1DeepSeek-V3.2-Exp — 2025-09-29V3.2-ExpDeepSeek-V3.2 and V3.2-Speciale — 2025-12-01V3.2R-SERIESDeepSeek-R1-Lite-Preview — 2024-11-20R1-LiteDeepSeek-R1 — 2025-01-20R1CODERDeepSeek-Coder-V2 — 2024-06-14V2
CURRENTSUPERSEDEDRETIREDLARGER SQUARE = WATERSHED
§ 02 — MODELS

Every DeepSeek model release.

15 model releases, newest first. “Current” is the newest release in its line; older ones are superseded, and retired ones have left DeepSeek’s apps or API. Where a model is in our benchmark tables, the last column links to its scores, price in AUD and Australian availability.

MODEL RELEASES15 RELEASES
DeepSeek model releases, newest first
RELEASEANNOUNCEDLINESTATUSBENCHMARKS
DeepSeek-V4.1-FlashFlashCURRENTDeepSeek V4.1 Flash →
DeepSeek-V4-Pro general availabilityProCURRENTDeepSeek V4 Pro →
DeepSeek-V4-Flash official releaseFlashRETIREDDeepSeek V4 Flash →
DeepSeek-V4 PreviewPro · FlashSUPERSEDED
DeepSeek-V3.2 and V3.2-SpecialeLLM & V-seriesRETIRED
DeepSeek-V3.2-ExpLLM & V-seriesSUPERSEDED
DeepSeek-V3.1LLM & V-seriesSUPERSEDED
Janus-ProJanusCURRENT
DeepSeek-R1WATERSHEDR-seriesSUPERSEDED
DeepSeek-V3WATERSHEDLLM & V-seriesSUPERSEDED
DeepSeek-R1-Lite-PreviewR-seriesSUPERSEDED
DeepSeek-V2.5LLM & V-seriesSUPERSEDED
DeepSeek-Coder-V2CoderSUPERSEDED
DeepSeek-V2LLM & V-seriesSUPERSEDED
DeepSeek LLM 7B and 67BLLM & V-seriesSUPERSEDED
§ 03 — TIMELINE

Every major release, year by year.

17 OF 17

20264 RELEASES

  1. MODEL

    DeepSeek-V4.1-Flash

    • LAUNCHED
    • CURRENT
    • V4 PRO AND FLASH

    DeepSeek-V4.1-Flash launches on the API as the smallest model in a new architecture family: a 552B-parameter causal encoder–decoder with 8B active parameters for input and 16B for output, native vision, and lower API prices. It replaces V4-Flash and the experimental V4-Flash-Vision-Exp.

    Why it mattered. DeepSeek reported it ahead of V4-Pro on benchmarks, with a KV cache needing a quarter of the memory, which cuts cache charges, often a large share of agent costs.

    AVAILABLE TO
    Open weights · API

  2. MODEL

    DeepSeek-V4-Pro general availability

    • GENERALLY AVAILABLE
    • CURRENT
    • V4 PRO AND FLASH

    The official DeepSeek-V4-Pro rolls out on the app and web (Expert Mode) and the API, with major agent upgrades, low/high/max reasoning effort and native OpenAI Responses API support. Peak and off-peak API pricing starts on 16 August, with off-peak rates 50% lower.

    Why it mattered. It replaced the preview with DeepSeek’s production flagship for agent work, and off-peak pricing gave API users a way to halve costs by scheduling jobs.

    AVAILABLE TO
    Open weights · Web · App · API

  3. MODEL

    DeepSeek-V4-Flash official release

    • PUBLIC BETA
    • RETIRED
    • V4 PRO AND FLASH

    The official DeepSeek-V4-Flash, a re-post-trained version of the preview with the same architecture and size, reaches the API in public beta with native Responses API support and open weights. The app and web models are unchanged.

    Why it mattered. DeepSeek reported it beating V4-Pro (Preview) on agent benchmarks despite far fewer active parameters, making the cheaper model the stronger choice for agent work.

    10 Sep 2026: Retired from the API when DeepSeek-V4.1-Flash launched; the deepseek-v4-flash model name temporarily routes to V4.1-Flash.

    AVAILABLE TO
    Open weights · API

  4. MODEL

    DeepSeek-V4 Preview

    • PREVIEW
    • SUPERSEDED
    • V4 PRO AND FLASH

    DeepSeek releases preview versions of DeepSeek-V4-Pro (1.6T total, 49B active parameters) and DeepSeek-V4-Flash (284B total, 13B active) as open weights, with a 1M token context window now the default across all DeepSeek services.

    Why it mattered. V4 split DeepSeek’s line-up into a flagship and a cheaper fast model, and made million-token context the standard for the app, web and API.

    AVAILABLE TO
    Open weights · Web · API

20256 RELEASES

  1. MODEL

    DeepSeek-V3.2 and V3.2-Speciale

    • LAUNCHED
    • RETIRED
    • HYBRID THINKING AND AGENTS

    DeepSeek-V3.2 succeeds V3.2-Exp on the app, web and API as DeepSeek’s first model to think during tool use. V3.2-Speciale, a maximum-reasoning variant that DeepSeek says reached gold-level results at IMO and IOI 2025, is offered on a temporary API endpoint.

    Why it mattered. Thinking inside tool calls made the everyday model better at agent work, and both models were released as open weights.

    24 Apr 2026: Replaced in the API by DeepSeek-V4-Flash, with the deepseek-chat and deepseek-reasoner names rerouted to it and then retired on 24 July 2026.

    AVAILABLE TO
    Open weights · Web · App · API

  2. MODEL

    DeepSeek-V3.2-Exp

    • EXPERIMENTAL
    • SUPERSEDED
    • HYBRID THINKING AND AGENTS

    DeepSeek-V3.2-Exp, an experimental model built on V3.1-Terminus, debuts DeepSeek Sparse Attention for more efficient long-context training and inference. It goes live on the app, web and API, and API prices drop by more than 50%.

    Why it mattered. DeepSeek reported performance on par with V3.1-Terminus, so API users got the same capability for less than half the price.

    AVAILABLE TO
    Open weights · Web · App · API

  3. MODEL

    DeepSeek-V3.1

    • LAUNCHED
    • SUPERSEDED
    • HYBRID THINKING AND AGENTS

    DeepSeek-V3.1 combines thinking and non-thinking modes in one model, switched with the DeepThink button, with stronger tool use and agent skills, 128K context, Anthropic API format support and open weights.

    Why it mattered. One hybrid model replaced the separate V3 and R1 lines, and DeepSeek called it its first step toward the agent era.

    AVAILABLE TO
    Open weights · Web · App · API

  4. MODEL

    Janus-Pro

    • OPEN WEIGHTS
    • CURRENT
    • R1 MOMENT

    DeepSeek releases Janus-Pro, an upgraded version of its Janus model that both understands images and generates them from text, as open weights in 1B and 7B sizes.

    Why it mattered. DeepSeek said it significantly improved both multimodal understanding and text-to-image instruction following over Janus, in a single open model.

    AVAILABLE TO
    Open weights

  5. MODEL

    DeepSeek-R1

    • WATERSHED
    • LAUNCHED
    • SUPERSEDED
    • R1 MOMENT

    DeepSeek releases R1, a reasoning model it reports on par with OpenAI o1 in maths, code and reasoning, under the MIT licence with six distilled smaller models. It is live on the web and API at $0.55/$2.19 per million input/output tokens.

    Why it mattered. An open-weights model that DeepSeek reported matching o1, with API outputs licensed for fine-tuning and distillation, showed that frontier-level reasoning could be released openly and priced low.

    AVAILABLE TO
    Open weights · Web · API

  6. PRODUCT

    DeepSeek app

    • LAUNCHED
    • R1 MOMENT

    DeepSeek releases its free mobile app on the App Store, Google Play and major Android stores, powered by DeepSeek-V3, with web search, Deep-Think mode, file upload and chat history synced across devices.

    Why it mattered. It brought DeepSeek chat to phones as a free official app, with no ads or in-app purchases, five days before R1 launched.

    AVAILABLE TO
    iOS · Android

20246 RELEASES

  1. MODEL

    DeepSeek-V3

    • WATERSHED
    • LAUNCHED
    • SUPERSEDED
    • R1 MOMENT

    DeepSeek-V3 launches as an open-weights 671B-parameter mixture-of-experts model with 37B active parameters, trained on 14.8 trillion tokens, running at 60 tokens a second. API prices stay at V2 levels until 8 February 2025, then $0.27/$1.10 per million input/output tokens.

    Why it mattered. DeepSeek said the release narrowed the gap between open and closed models, and V3 became the base for DeepSeek’s models through 2025, including R1.

    AVAILABLE TO
    Open weights · Web · API

  3. MODEL

    DeepSeek-R1-Lite-Preview

    • PREVIEW
    • SUPERSEDED
    • R1 MOMENT

    DeepSeek launches R1-Lite-Preview, its first reasoning model, on chat.deepseek.com, showing its thought process in real time and reporting o1-preview-level results on the AIME and MATH benchmarks.

    Why it mattered. It was DeepSeek’s first step into o1-style reasoning and previewed the R1 line, with open-source models and API access promised to follow.

    AVAILABLE TO
    Web

  4. MODEL

    DeepSeek-V2.5

    • LAUNCHED
    • SUPERSEDED
    • OPEN-MODEL FOUNDATIONS

    DeepSeek-V2.5 merges the DeepSeek-V2 chat model and DeepSeek-Coder-V2 into one model, with better writing and instruction following, on the web, the API and as open weights.

    Why it mattered. One model now served both the deepseek-chat and deepseek-coder endpoints, so users no longer had to choose between a chat model and a coding model.

    AVAILABLE TO
    Open weights · Web · API

  5. MODEL

    DeepSeek-Coder-V2

    • LAUNCHED
    • SUPERSEDED
    • OPEN-MODEL FOUNDATIONS

    DeepSeek upgrades its deepseek-coder API model to DeepSeek-Coder-V2 and releases open-weights versions with 16B and 236B mixture-of-experts parameters.

    Why it mattered. DeepSeek said it reached the level of GPT-4-Turbo-0409 in code generation, understanding, debugging and completion, putting an open coding model alongside the leading closed ones.

    AVAILABLE TO
    Open weights · Web · API

  6. MODEL

    DeepSeek-V2

    • LAUNCHED
    • SUPERSEDED
    • OPEN-MODEL FOUNDATIONS

    DeepSeek-V2 launches as a 236B-parameter mixture-of-experts model with 21B active parameters and 128K context, available as open weights, on chat.deepseek.com and on DeepSeek’s OpenAI-compatible API.

    Why it mattered. DeepSeek said it cut training costs by 42.5% and the KV cache by 93.3% compared with DeepSeek 67B, and it offered API access at what it called an unbeatable price.

    AVAILABLE TO
    Open weights · Web · API

20231 RELEASE

  1. MODEL

    DeepSeek LLM 7B and 67B

    • OPEN WEIGHTS
    • SUPERSEDED
    • OPEN-MODEL FOUNDATIONS

    DeepSeek releases DeepSeek LLM in 7B and 67B sizes, base and chat, trained from scratch on 2 trillion English and Chinese tokens, with commercial use permitted under its model licence.

    Why it mattered. It was DeepSeek’s first general-purpose model family, and DeepSeek reported the 67B base outperforming Llama 2 70B on reasoning, coding, maths and Chinese.

    AVAILABLE TO
    Open weights
§ 04 — ERAS

4 eras of DeepSeek.

  1. ERA 01NOVEMBER 2023 – SEPTEMBER 2024

    Open-model foundations

    DeepSeek LLM, the low-cost DeepSeek-V2 mixture-of-experts model, Coder-V2 and the merged V2.5 establish DeepSeek as an open-weights lab with a cheap API.

    4 releases, from DeepSeek LLM 7B and 67B →
  2. ERA 02NOVEMBER 2024 – JANUARY 2025

    R1 moment

    R1-Lite-Preview, Internet Search, DeepSeek-V3, the mobile app, R1 and Janus-Pro arrive within about ten weeks, turning DeepSeek into a consumer product built on openly released frontier-level models.

    6 releases, from DeepSeek-R1-Lite-Preview →
  3. ERA 03AUGUST 2025 – DECEMBER 2025

    Hybrid thinking and agents

    V3.1 merges thinking and non-thinking modes into one model, V3.2-Exp adds sparse attention with a 50% API price cut, and V3.2 brings thinking into tool use.

    3 releases, from DeepSeek-V3.1 →
  4. ERA 04APRIL 2026 – NOW

    V4 Pro and Flash

    The V4 preview makes 1M context standard, official V4-Flash and V4-Pro releases follow, and V4.1-Flash debuts a new architecture with native vision.

    4 releases, from DeepSeek-V4 Preview →
§ 05 — QUESTIONS

DeepSeek releases, answered.

As of our 23 Sep 2026 review, DeepSeek-V4.1-Flash, announced by DeepSeek on 10 Sep 2026. DeepSeek-V4.1-Flash launches on the API as the smallest model in a new architecture family: a 552B-parameter causal encoder–decoder with 8B active parameters for input and 16B for output, native vision, and lower API prices. It replaces V4-Flash and the experimental V4-Flash-Vision-Exp.

29 Nov 2023: DeepSeek LLM 7B and 67B. DeepSeek releases DeepSeek LLM in 7B and 67B sizes, base and chat, trained from scratch on 2 trillion English and Chinese tokens, with commercial use permitted under its model licence.

As of 23 Sep 2026, the newest release in each line is: Janus-Pro (27 Jan 2025), DeepSeek-V4-Pro general availability (13 Aug 2026) and DeepSeek-V4.1-Flash (10 Sep 2026). Older models in those lines have been superseded, and some have been retired from DeepSeek’s apps.

17 on this timeline: 15 model releases and 2 product, platform, agent and plan launches. It leaves out incremental updates, settings and feature drops — only releases that changed what DeepSeek could do or who could use it.

DeepSeek V4 Pro: it can be called from Microsoft Azure in Australia East, but only through global routing, so requests may be processed outside Australia; its weights are public, so it can also be self-hosted in Australia. DeepSeek V4.1 Flash: Its weights are public, so it can be hosted in Australia on your own infrastructure or an Australian cloud region. Our benchmark pages track Australian cloud availability for each model weekly.

Dates are the day DeepSeek announced each release, taken from DeepSeek’s own site and documentation.

Major releases only: new models, products, plans and capability launches. Retirements are noted on the model they retired.

Checked 23 Sep 2026 · How the current models compare →

MAKE SENSE OF THE NEXT RELEASE

Choose AI improvements around your business needs.

DISCUSS YOUR AI OPTIONSAI SUPPORT & MAINTENANCE