DeepSeek is a Chinese artificial intelligence company founded in Hangzhou in 2023. Its first public model family, DeepSeek Coder, arrived on November 2, 2023. The company then expanded into general language models, mathematical reasoning, formal proof, vision, document recognition, and dedicated reasoning models.
DeepSeek’s newest model family is DeepSeek-V4 Preview, released on April 24, 2026. The family includes DeepSeek-V4-Pro and DeepSeek-V4-Flash. Both models support thinking and non-thinking modes, a one-million-token context window, and downloadable weights.
Between DeepSeek Coder and V4 Preview, the company repeatedly folded ideas from its specialist research into its main model line. The dates below show how that progression unfolded, first in a quick reference and then in the full year-by-year timeline.
Last Updated: July 10, 2026
Key DeepSeek Dates
| Milestone | Date and details |
|---|---|
| DeepSeek founded | 2023 Liang Wenfeng founded the company with financial backing from High-Flyer. |
| First public model | November 2, 2023 DeepSeek Coder was the company’s first public model family. |
| DeepSeek-V3 | December 26, 2024 V3 became the foundation for R1 and powered the DeepSeek App at launch. |
| DeepSeek App | January 15, 2025 The official App launched with DeepSeek-V3. |
| DeepSeek-R1 | January 20, 2025 R1 arrived five days after the App launched. |
| DeepSeek-V4 Preview | April 24, 2026 The latest model family includes V4-Pro and V4-Flash. |
DeepSeek Release History
| Release | Date and key change |
|---|---|
| DeepSeek-V4 Preview | April 24, 2026 Current preview family with Pro and Flash models, dual reasoning modes, and a 1M context window. |
| DeepSeek-OCR-2 | January 27, 2026 Second document-understanding model based on visual causal flow. |
| DeepSeek-V3.2 | December 1, 2025 Added thinking during tool use. Open weights remain available after V4 replaced it on the official API. |
| DeepSeekMath-V2 | November 27, 2025 Mathematical reasoning model built around verification and proof-oriented training. |
| DeepSeek-OCR | October 20, 2025 Research model for document recognition and visual-text compression. |
| DeepSeek-V3.2-Exp | September 29, 2025 Experimental release that introduced DeepSeek Sparse Attention. |
| DeepSeek-V3.1-Terminus | September 22, 2025 V3.1 update for language consistency and agent performance. |
| DeepSeek-V3.1 | August 21, 2025 Combined thinking and non-thinking modes in one model and improved tool use. |
| DeepSeek-R1-0528 | May 28, 2025 Updated R1 with stronger reasoning, fewer hallucinations, JSON output, and function calling. |
| DeepSeek-Prover-V2 | April 30, 2025 Formal theorem-proving models released in 7B and 671B sizes. |
| DeepSeek-V3-0324 | March 24, 2025 Improved reasoning, coding, writing, and tool use. Model weights moved to the MIT License. |
| Janus-Pro | January 27, 2025 Updated unified multimodal models for image understanding and generation. |
| DeepSeek-R1 | January 20, 2025 Reasoning family with R1-Zero, R1, and six distilled models. |
| DeepSeek App | January 15, 2025 Official iOS and Android App, initially powered by V3. |
| DeepSeek-V3 | December 26, 2024 671B-parameter MoE model with 37B active parameters per token. |
| DeepSeek-VL2 | December 13, 2024 Second vision-language family using a mixture-of-experts design. |
| DeepSeek-R1-Lite-Preview | November 20, 2024 Hosted reasoning preview that preceded the open R1 release. |
| Janus | October 2024 Unified multimodal research family for image understanding and generation. |
| DeepSeek-V2.5 | September 5, 2024 Merged the general abilities of V2 Chat with the coding abilities of Coder-V2. |
| DeepSeek-Coder-V2 | June 17, 2024 MoE coding family with support for 338 programming languages and 128K context. |
| DeepSeek-Prover | May 2024 First DeepSeek research family for formal theorem proving in Lean. |
| DeepSeek-V2 | May 6, 2024 Introduced the general MoE architecture and Multi-head Latent Attention used by later flagships. |
| DeepSeek-VL | March 2024 First DeepSeek vision-language family. |
| DeepSeekMath | February 2024 Math-focused 7B models released as Base, Instruct, and RL checkpoints. |
| DeepSeekMoE | January 11, 2024 Introduced fine-grained expert routing and shared-expert isolation. |
| DeepSeek LLM | November 29, 2023 First general-purpose DeepSeek language models, released in 7B and 67B sizes. |
| DeepSeek Coder | November 2, 2023 First public DeepSeek model family, built for code generation and completion. |
Current DeepSeek Model Lineup
DeepSeek-V4-Pro is the larger model in the V4 Preview family. It has 1.6 trillion total parameters and activates 49 billion for each token. DeepSeek recommends it for difficult reasoning, coding, knowledge, and agent tasks. DeepSeek-V4-Flash is the faster option, with 284 billion total parameters and 13 billion active per token.
Both V4 models offer thinking and non-thinking modes within a one-million-token context window. The web product labels the model choices Expert Mode and Instant Mode. API users select deepseek-v4-pro or deepseek-v4-flash.
DeepSeek-V3.2 was the previous general flagship. Its open weights remain available, but V4 has replaced it on DeepSeek’s main API routes. V3.2 added reasoning during tool use and completed the architecture first tested in V3.2-Exp.
The API identifiers deepseek-chat and deepseek-reasoner are compatibility names rather than permanent model versions. As of July 10, 2026, they route to the non-thinking and thinking modes of V4-Flash. DeepSeek has scheduled both identifiers for retirement after July 24, 2026, so new integrations should use the V4 model names directly.
DeepSeek Timeline
Table Of Contents
2026: OCR-2 and DeepSeek-V4 Preview
| Date | Release or Milestone and Why It Mattered |
|---|---|
| April 24, 2026 | DeepSeek-V4 Preview: DeepSeek released V4-Pro and V4-Flash through its web service, API, and open-weight repositories. Pro handles more demanding work, while Flash favors speed. Both have a 1M context window and selectable thinking modes. |
| January 27, 2026 | DeepSeek-OCR-2: The second OCR research release replaced the first model’s standard visual processing order with visual causal flow. The change was designed to compress long documents into more efficient visual representations. |
2025: R1, Hybrid Reasoning, Agents, and Specialist Models
| Date | Release or Milestone and Why It Mattered |
|---|---|
| December 1, 2025 | DeepSeek-V3.2: V3.2 arrived in the App, web service, API, and open-weight repositories. It could use tools in both thinking and non-thinking modes. A temporary V3.2-Speciale endpoint offered heavier reasoning until December 15. |
| November 27, 2025 | DeepSeekMath-V2: DeepSeek returned to specialist mathematical reasoning with a much larger model centered on verifiable generation. The release kept the separate math research line active after R1 brought advanced math into a general reasoning model. |
| October 20, 2025 | DeepSeek-OCR: This 3B document model represented text-heavy pages with compact visual tokens. It established OCR as a separate research branch rather than a feature inside the main chat model. |
| September 29, 2025 | DeepSeek-V3.2-Exp: The experimental model introduced DeepSeek Sparse Attention to reduce the cost of long-context processing. It served as the public architecture test for V3.2. |
| September 22, 2025 | DeepSeek-V3.1-Terminus: This corrective update reduced Chinese-English mixing and stray characters. It also improved code-agent and search-agent behavior without starting a new model generation. |
| August 21, 2025 | DeepSeek-V3.1: V3.1 put thinking and non-thinking behavior in one model and improved tool use for multi-step agent tasks. The V series was no longer limited to conventional chat and code completion. |
| May 28, 2025 | DeepSeek-R1-0528: The update improved mathematics, coding, and front-end generation. JSON output and function calling also made R1 easier to use in software integrations. |
| April 30, 2025 | DeepSeek-Prover-V2: DeepSeek released theorem-proving models in 7B and 671B sizes. The larger version brought the formal-proof branch onto the V3-scale architecture. |
| March 24, 2025 | DeepSeek-V3-0324: The update improved reasoning, coding, Chinese writing, search, and function calling. DeepSeek released the revised weights under the MIT License. |
| January 27, 2025 | Janus-Pro: Janus gained better image understanding and generation. It remained a downloadable research family, not a multimodal mode in the main DeepSeek App. |
| January 20, 2025 | DeepSeek-R1: DeepSeek released R1-Zero, R1, and six smaller distilled models. Its reinforcement-learning approach improved step-by-step reasoning and drew attention far beyond the developer community. |
| January 15, 2025 | DeepSeek App: The official App launched on iOS and Android with V3 underneath. R1 followed five days later as a separate model release. |
2024: MoE Research, V2, V3, and the First Reasoning Preview
| Date | Release or Milestone and Why It Mattered |
|---|---|
| December 26, 2024 | DeepSeek-V3: V3 expanded the company’s MoE design to 671 billion total parameters, with 37 billion active per token. It became the base for R1 and powered the DeepSeek App at launch. |
| December 13, 2024 | DeepSeek-VL2: The second vision-language family adopted an MoE design and came in Tiny, Small, and full-size variants. DeepSeek kept it separate from the text-only V3 family. |
| November 20, 2024 | DeepSeek-R1-Lite-Preview: DeepSeek put an early reasoning model on its web and API services without releasing the weights. The preview let users try the approach that later became R1. |
| October 2024 | Janus: Janus assigned different visual encoding paths to image understanding and generation but shared one language-model core. That split allowed one system to handle both jobs without forcing them through the same visual representation. |
| September 5, 2024 | DeepSeek-V2.5: DeepSeek combined V2 Chat and Coder-V2. General conversation and code work no longer needed separate official API model families. |
| June 17, 2024 | DeepSeek-Coder-V2: The second Coder generation adopted an MoE architecture, increased language coverage from 86 to 338 programming languages, and expanded context from 16K to 128K. |
| May 2024 | DeepSeek-Prover: The first Prover models targeted formal theorem proving in Lean. Unlike DeepSeekMath’s natural-language answers, a Lean proof assistant could check these proofs directly. |
| May 6, 2024 | DeepSeek-V2: V2 introduced Multi-head Latent Attention and a larger MoE design. Both efficiency ideas remained central to V3, R1, and later V-series models. |
| March 2024 | DeepSeek-VL: DeepSeek’s first vision-language models could interpret images, charts, and text within images. They opened a multimodal research line alongside the text models. |
| February 2024 | DeepSeekMath: DeepSeek published Base, Instruct, and reinforcement-learning versions of a 7B math model. Its Group Relative Policy Optimization method later became part of the R1 training work. |
| January 11, 2024 | DeepSeekMoE: The research model divided experts into smaller units and isolated shared knowledge in dedicated experts. DeepSeek carried this sparse design into its later flagship models. |
2023: DeepSeek Begins with Code and General Language Models
| Date | Release or Milestone and Why It Mattered |
|---|---|
| November 29, 2023 | DeepSeek LLM: DeepSeek released general-purpose 7B and 67B models in Base and Chat forms. They were the company’s first public models for language tasks beyond programming. |
| November 2, 2023 | DeepSeek Coder: DeepSeek’s first public model family handled code completion, generation, and repository-level context. Coding was the company’s starting point, not a later addition to a chat model. |
| 2023 | Company founding: Liang Wenfeng founded DeepSeek in Hangzhou with financial backing from High-Flyer. The new company pursued artificial general intelligence research rather than extending High-Flyer’s trading business. |
DeepSeek Model Families Explained
V Series
The V series is DeepSeek’s main general-purpose model line. V2 introduced the MoE and attention architecture that shaped later releases. V2.5 combined general and coding abilities, V3 increased the scale, V3.1 merged thinking with regular response modes, and V3.2 brought reasoning into tool use. V4 now divides the main line into Pro and Flash models.
R Series
The R series focuses on reasoning. R1-Lite-Preview was the first hosted preview. R1-Zero tested reinforcement learning without supervised fine-tuning as the first post-training stage, while R1 added cold-start data to improve readability and consistency. The six distilled R1 models transferred parts of that reasoning behavior into smaller Qwen- and Llama-based models.
Coder, Math, and Prover
DeepSeek Coder and Coder-V2 were built for programming. DeepSeekMath concentrated on competition-style mathematical problems expressed in natural language. DeepSeek-Prover targeted formal proofs that a Lean proof assistant could check. These branches influenced the general models even when their names disappeared from the main API.
VL, Janus, and OCR
DeepSeek-VL and VL2 handle vision-language understanding. Janus combines image understanding with image generation, but uses different visual encoding paths for the two jobs. DeepSeek-OCR treats document images as compressed carriers of text and layout. These are downloadable research families, not names for the current text-only V4 API models.
What the Model Suffixes Mean
| Suffix | Meaning |
|---|---|
| Base | A pretrained checkpoint before instruction tuning for conversational use. |
| Chat or Instruct | A model tuned to follow instructions and respond in a conversational format. |
| Distill | A smaller model trained with outputs or reasoning traces from a larger model. |
| Lite | A smaller model or preview intended to use fewer computing resources. |
| Preview | An early public release that is not presented as the final version. |
| Exp | An experimental release used to test a new architecture or serving method. |
| Terminus | The final corrective update in the V3.1 line before the V3.2 experimental branch. |
Five Turning Points in DeepSeek’s History
1. The First Releases Were Specialist Models
DeepSeek did not begin with a consumer chatbot. It released a code model first, then a general language model, followed by separate research families for MoE routing, mathematics, and vision. Each branch tested an idea that could later move into a larger general model.
The specialist work also fed later releases. DeepSeekMath explored reinforcement-learning methods used in R1. DeepSeekMoE supplied the sparse expert design behind V2 and V3. Coder and Math data helped train Coder-V2 and the general models that followed.
2. V2 Established the Efficiency Architecture
V2 was the first general DeepSeek model built around the architecture now associated with the company. Its mixture-of-experts system activated only part of the full parameter set for each token. Multi-head Latent Attention reduced the memory required for the key-value cache during inference.
Those choices carried into later releases. V3 used a much larger MoE design, R1 inherited the V3 base architecture, and V4 continued the effort to lower the cost of long-context inference. V2 is the architectural starting point for the later flagships.
3. V3 Became the General Foundation
V3 expanded the MoE design to 671 billion total parameters while activating 37 billion for each token. It handled general writing, coding, mathematics, and other text tasks through one model. The DeepSeek App used V3 at launch, and the official deepseek-chat API identifier pointed to V3 at the time.
V3 also became the base for R1. R1 was not a separate architecture built from the ground up. It applied a reasoning-focused post-training process to the capabilities and efficiency of the V3 foundation.
4. R1 Took DeepSeek Beyond Its Developer Audience
R1 made DeepSeek widely known in January 2025. The release combined a large reasoning model, an experimental R1-Zero checkpoint, and six smaller distilled models. Developers could inspect and run the weights, while ordinary users could access reasoning through the DeepSeek service.
The App had launched five days earlier with V3. R1 expanded what users could do inside the existing product, but it was not the App’s launch model. The later R1-0528 update added JSON output and function calling for software integrations.
5. V3.1 and V4 Merged Reasoning with Agent Work
V3.1 reduced the separation between a normal chat model and a reasoning model. One model could answer in thinking or non-thinking mode, and the API identifiers selected the mode. V3.2 extended that design by allowing the model to reason while using tools.
V4 kept both modes and gave Pro and Flash a 1M context window. DeepSeek also treated coding, tool use, long inputs, and multi-step agent tasks as core workloads rather than secondary chat features.
DeepSeek Timeline FAQs
When was DeepSeek founded?
DeepSeek was founded in Hangzhou, China, in 2023 by Liang Wenfeng. High-Flyer, the quantitative investment firm he co-founded, provided financial backing. No single founding day is consistently documented because the research lab’s creation and its later corporate separation occurred at different stages.
What was the first DeepSeek model?
DeepSeek Coder was the first public DeepSeek model family. It was released on November 2, 2023, with models for code completion and instruction-based programming tasks. DeepSeek LLM followed on November 29 as the company’s first general-purpose language model family.
When was DeepSeek-R1 released?
DeepSeek-R1 was released on January 20, 2025. The release included DeepSeek-R1-Zero, the main R1 model, and six distilled models based on Qwen and Llama architectures. DeepSeek updated the reasoning family with R1-0528 on May 28, 2025.
Did the DeepSeek App launch with R1?
No. The official DeepSeek App launched on January 15, 2025, and was initially powered by DeepSeek-V3. R1 arrived on January 20 and was then made available through DeepSeek’s web product, App, and API. The two events happened in the same week but were separate releases.
What is the latest DeepSeek model?
DeepSeek-V4 Preview is the company’s latest model family. DeepSeek released it on April 24, 2026, as V4-Pro and V4-Flash. The Preview label remains part of the official name; DeepSeek has not presented it as the final V4 generation.
What is the difference between DeepSeek V models and R models?
The V series is DeepSeek’s main general-purpose line. It covers ordinary chat, writing, coding, tool use, and other text tasks. The R series was created for deliberate reasoning, especially mathematics, coding, and multi-step problems. Starting with V3.1, the V series also gained thinking modes, so the boundary is less rigid than it was when R1 launched.
Is DeepSeek open source?
Many DeepSeek models have downloadable weights, technical reports, and permissive licenses. R1 and several recent V-series releases use the MIT License.
What do deepseek-chat and deepseek-reasoner mean?
deepseek-chat and deepseek-reasoner are API compatibility identifiers. Their underlying models changed when DeepSeek upgraded its service. On July 10, 2026, both point to V4-Flash: the first selects non-thinking mode and the second selects thinking mode. DeepSeek plans to retire both identifiers after July 24, 2026.
