AI Radar

DeepSeek models

DeepSeek (Hangzhou DeepSeek Artificial Intelligence Co., Ltd.) is a Chinese AI company that says it was founded in 2023. It describes itself as a research team focused on fundamental model research with an open-source approach, and says it releases model weights under the MIT License with a technical report for each model. It offers its models through a web chat, mobile apps and an API platform.

Chronology

  1. 2026

    6 models
    1. Released
      DeepSeek Flash · Language, Multimodal, Reasoning DeepSeek-V4.1-Flash

      DeepSeek-V4.1-Flash is a Mixture-of-Experts model with 552B backbone parameters that takes text and images and generates text, released on 2026-09-10 in the DeepSeek API as deepseek-flash and with…

      • Processes images natively together with text, trained jointly from the start of pre-training; the previous Flash model accepted text only.
      • Switches to a Causal Encoder-Decoder design with 8B active parameters for input and 16B for output, plus CSA2 sparse attention and Engram conditional memory.
      Primary source Available · 7 news items
    2. Released
      DeepSeek Flash · Language, Multimodal, Reasoning DeepSeek-V4-Flash-Vision-Exp

      DeepSeek-V4-Flash-Vision-Exp is an experimental model that DeepSeek calls its first multimodal model in the DeepSeek-V4 family, built on the V4-Flash architecture with added visual modules. It…

      • Adds image input next to text; images can be sent as base64, as external URLs or through the new Files API.
      • Adds visual modules to the V4-Flash architecture, followed by continued training.
      Primary source Available · no news linked
    3. Released
      DeepSeek Pro · Language, Reasoning DeepSeek-V4-Pro-0813

      DeepSeek-V4-Pro-0813 is the general-availability release of DeepSeek-V4-Pro, rolled out on 2026-08-13 in the DeepSeek app, web chat (Expert Mode) and API under the unchanged name deepseek-v4-pro…

      • Adds an attached DSpark speculative decoding module on top of the preview's model structure.
      • Natively supports the OpenAI Responses API format, adapted for Codex with a one-click setup script.
      Primary source Available · no news linked
    4. Released
      DeepSeek Flash · Language, Reasoning DeepSeek-V4-Flash-0731

      DeepSeek-V4-Flash-0731 is the official, non-preview release of DeepSeek-V4-Flash, with open weights under the MIT License, served in the DeepSeek API from 2026-07-31 under the unchanged name…

      • Ships with an attached DSpark speculative decoding module; DeepSeek says architecture and size are otherwise unchanged and only post-training was redone.
      • Reasoning effort is set through a reasoning_effort parameter with three levels: low, high and max.
      Primary source Available · no news linked
    5. Released
      DeepSeek Flash · Language, Reasoning · Milestone DeepSeek-V4-Flash

      DeepSeek-V4-Flash is an open-weights Mixture-of-Experts language model with 284B total and 13B active parameters, the smaller model of the DeepSeek-V4 preview released on 2026-04-24 alongside…

      • Uses the V4 hybrid attention design (Compressed Sparse Attention plus Heavily Compressed Attention), Manifold-Constrained Hyper-Connections and the Muon optimizer.
      • 284B total and 13B activated parameters, against 671B total and 37B activated for DeepSeek-V3.2-Base in the card's base-model table.
      Primary source Available · no news linked
    6. Released
      DeepSeek Pro · Language, Reasoning · Milestone DeepSeek-V4-Pro

      DeepSeek-V4-Pro is an open-weights Mixture-of-Experts language model with 1.6T total and 49B active parameters, released on 2026-04-24 as the preview of the DeepSeek-V4 series alongside…

      • Introduces hybrid attention (Compressed Sparse Attention plus Heavily Compressed Attention), Manifold-Constrained Hyper-Connections and training with the Muon optimizer.
      • At a 1M-token context it needs about 27% of the single-token inference FLOPs and 10% of the KV cache of DeepSeek-V3.2, according to DeepSeek.
      Primary source Available · no news linked

  2. 2025

    9 models
    1. Released
      DeepSeek-V · Language, Reasoning DeepSeek-V3.2

      DeepSeek-V3.2 is a text-only language model in DeepSeek's V line with thinking and non-thinking modes, released on 1 December 2025 in the app, web chat and API, with open weights under the MIT…

      • Integrates thinking directly into tool use, which DeepSeek says is a first for its models, and supports tool use in both thinking and non-thinking modes.
      • Trained with synthesised agent data that DeepSeek says covers more than 1,800 environments and 85,000 complex instructions.
      Primary source Available · no news linked
    2. Released
      DeepSeek-V · Language, Reasoning DeepSeek-V3.2-Speciale

      DeepSeek-V3.2-Speciale is a high-compute variant of DeepSeek-V3.2 designed for deep reasoning tasks, announced with it on 1 December 2025 and released with open weights under the MIT License. It…

      • Trained only on reasoning data with a reduced length penalty in reinforcement learning, plus the DeepSeekMath-V2 dataset and reward method for mathematical proofs.
      • Does not support tool calling.
      Primary source Available · no news linked
    3. Released
      DeepSeek-V · Language, Reasoning DeepSeek-V3.2-Exp

      DeepSeek-V3.2-Exp is an experimental language model in DeepSeek's V line, released on 29 September 2025 in the app, web chat and API, with open weights under the MIT License. It adds DeepSeek Sparse…

      • Adds DeepSeek Sparse Attention, a fine-grained sparse attention mechanism with a lightning indexer; DeepSeek says this is the only architectural change.
      • DeepSeek says sparse attention makes long-context training and inference more efficient; it cut API prices by more than 50 percent with this release.
      Primary source Available · no news linked
    4. Released
      DeepSeek-V · Language, Reasoning DeepSeek-V3.1-Terminus

      DeepSeek-V3.1-Terminus is an update of DeepSeek-V3.1, released on 22 September 2025 in the DeepSeek app, web chat and API, with open weights under the MIT License. DeepSeek says it keeps the…

      • Less mixing of Chinese and English and fewer occasional abnormal characters in output, according to DeepSeek.
      • DeepSeek reports better Code Agent and Search Agent performance; the search-agent template and tool set were updated.
      Primary source Available · no news linked
    5. Released
      DeepSeek-V · Language, Reasoning DeepSeek-V3.1

      DeepSeek-V3.1 is a hybrid language model in DeepSeek's V line that offers a thinking and a non-thinking mode in one model, launched on 21 August 2025 in DeepSeek's web chat and API. Its weights and…

      • One hybrid model serves both a thinking and a non-thinking mode; in the API these modes had been served by separate models, DeepSeek-V3-0324 and DeepSeek-R1-0528.
      • DeepSeek says post-training improved tool use and performance on agent tasks.
      Primary source Available · no news linked
    6. Released
      DeepSeek-R · Language, Reasoning DeepSeek-R1-0528

      DeepSeek-R1-0528 is a minor version upgrade of DeepSeek-R1, served behind deepseek-reasoner in the DeepSeek API from 28 May 2025, in the web chat and as open weights under the MIT License. DeepSeek…

      • JSON output and function calling support are listed as part of the upgrade.
      • Deeper reasoning, which DeepSeek attributes to more compute and algorithmic optimisation in post-training; complex tasks may use more tokens.
      Primary source Available · no news linked
    7. Released
      DeepSeek-V · Language DeepSeek-V3-0324

      DeepSeek-V3-0324 is an updated version of DeepSeek-V3 with the same model structure, served behind deepseek-chat in the DeepSeek API from 24 March 2025, in the web chat and as open weights under the…

      • Weights released under the MIT License instead of the DeepSeek Model License used for DeepSeek-V3.
      • Improved reasoning, according to DeepSeek, with the model structure unchanged.
      Primary source Available · no news linked
    8. Released
      DeepSeek-R · Language, Reasoning · Milestone DeepSeek-R1

      DeepSeek-R1 is a reasoning language model built on DeepSeek-V3-Base, released on 20 January 2025 in the DeepSeek web chat (DeepThink), in the API as deepseek-reasoner and as open weights under the…

      • Offered in the API and as open weights, while DeepSeek-R1-Lite-Preview was available only in the DeepSeek web chat at launch.
      • Code and weights under the MIT License, and DeepSeek allows the weights and API outputs to be used for fine-tuning and distillation.
      Primary source Available · no news linked
    9. Released
      DeepSeek-R · Language, Reasoning DeepSeek-R1-Zero

      DeepSeek-R1-Zero is one of DeepSeek's first-generation reasoning models, trained with large-scale reinforcement learning directly on DeepSeek-V3-Base without supervised fine-tuning first, and…

      • Reasoning is learned through large-scale reinforcement learning (GRPO) applied directly to DeepSeek-V3-Base, with no supervised fine-tuning stage first.
      Primary source Available · no news linked

  3. 2024

    7 models
    1. Released
      DeepSeek-V · Language · Milestone DeepSeek-V3

      DeepSeek-V3 is a mixture-of-experts language model with 671B total and 37B activated parameters, offered in DeepSeek's web chat, through its API and as open weights together with the base model…

      • Total parameters rise to 671B with 37B activated per token, from the 236B total and 21B activated that DeepSeek lists for DeepSeek-V2 and V2.5.
      • Adds an auxiliary-loss-free load-balancing strategy and a multi-token prediction training objective to the DeepSeek-V2 architecture.
      Primary source Available · no news linked
    2. Released
      DeepSeek-V · Language DeepSeek-V2.5-1210

      DeepSeek-V2.5-1210 is an upgraded version of DeepSeek-V2.5, released on 10 December 2024 behind deepseek-chat in the DeepSeek API and as open weights on Hugging Face. DeepSeek presented it as the…

      • DeepSeek reports improvements in maths, coding, writing and reasoning over DeepSeek-V2.5.
      • DeepSeek says the update improves the user experience for file upload and web page summarisation.
      Primary source Available · no news linked
    3. Released
      DeepSeek-R · Language, Reasoning DeepSeek-R1-Lite-Preview

      DeepSeek-R1-Lite-Preview is a preview of a DeepSeek reasoning model that went live in the DeepSeek web chat on 20 November 2024. DeepSeek highlighted that it shows its thought process in real time…

      Primary source Status unknown · no news linked
    4. Released
      DeepSeek-V · Language DeepSeek-V2.5

      DeepSeek-V2.5 is a language model that DeepSeek made by merging DeepSeek-V2-0628 with its code model DeepSeek-Coder-V2-0724, to combine general conversation and coding in one model. It was released…

      • Merges the general chat model DeepSeek-V2-0628 with the code model DeepSeek-Coder-V2-0724 into one model.
      • One model now serves both the deepseek-chat and deepseek-coder API models, which DeepSeek kept for backward compatibility.
      Primary source Available · no news linked
    5. Released
      DeepSeek-V · Language DeepSeek-V2-Chat-0628

      DeepSeek-V2-Chat-0628 is a revised DeepSeek-V2 chat model that went live in the DeepSeek API on 2024-06-28 as DeepSeek-V2-0628 and was later published as open weights under the DeepSeek Model…

      • Built on the Coder-V2 base model instead of the DeepSeek-V2 base; DeepSeek says the switch was meant to improve code generation and reasoning.
      • Instruction following in the system prompt improved, which the model card says helps tasks such as immersive translation and retrieval-augmented generation.
      Primary source Available · no news linked
    6. Released
      DeepSeek-V · Language · Milestone DeepSeek-V2

      DeepSeek-V2 is a mixture-of-experts language model from DeepSeek with 236B total parameters, 21B of them activated per token, and a 128K-token context, released as open weights and offered on…

      • Moves from a dense transformer to a mixture-of-experts design (DeepSeekMoE) and adds the new Multi-head Latent Attention.
      • Context length of 128K tokens, up from the 4,096-token sequence length of DeepSeek LLM.
      Primary source Available · no news linked
    7. Released
      DeepSeek · Language DeepSeekMoE 16B

      DeepSeekMoE 16B is an open-weights mixture-of-experts language model from DeepSeek with 16.4B total and about 2.8B activated parameters, released as a Base and a Chat checkpoint with the DeepSeekMoE…

      • Uses a sparse mixture-of-experts architecture (DeepSeekMoE) instead of the dense transformer of DeepSeek LLM 7B.
      • DeepSeek reports results comparable with its dense DeepSeek 7B, trained on the same corpus, with about 40% of the computation.
      Primary source Available · no news linked

  4. 2023

    1 model
    1. Released
      DeepSeek · Language · Milestone DeepSeek LLM

      DeepSeek LLM is a general-purpose language model series from DeepSeek, released as open weights in 7B and 67B sizes, each as a Base and a Chat model, under the DeepSeek Model License, which permits…

      Primary source Available · no news linked

Lineage

  1. DeepSeek LLM
    Successors
    DeepSeek-V2
  2. DeepSeekMoE 16B

    No recorded relations.

  3. DeepSeek-V2
    Predecessors
    DeepSeek LLM
    Variants and derived
    DeepSeek-V2-Chat-0628 · revision
  4. DeepSeek-V2-Chat-0628
    Based on
    DeepSeek-V2 · revision
    Variants and derived
    DeepSeek-V2.5 · derived: merge
    Siblings
    DeepSeek-V2 (base model) · inline variant of DeepSeek-V2; DeepSeek-V2-Chat · inline variant of DeepSeek-V2; DeepSeek-V2-Lite · inline variant of DeepSeek-V2; DeepSeek-V2-Lite-Chat · inline variant of DeepSeek-V2
  5. DeepSeek-V2.5
    Based on
    DeepSeek-V2-Chat-0628 · derived: merge
    Variants and derived
    DeepSeek-V2.5-1210 · revision
  6. DeepSeek-R1-Lite-Preview
    Successors
    DeepSeek-R1
  7. DeepSeek-V2.5-1210
    Successors
    DeepSeek-V3
    Based on
    DeepSeek-V2.5 · revision
  8. DeepSeek-V3
    Predecessors
    DeepSeek-V2.5-1210
    Variants and derived
    DeepSeek-R1 · derived: fine tune; DeepSeek-R1-Zero · derived: other; DeepSeek-V3-0324 · revision; DeepSeek-V3.1 · derived: continued pretraining
  9. DeepSeek-R1
    Based on
    DeepSeek-V3 · derived: fine tune
    Variants and derived
    DeepSeek-R1-0528 · revision; MAI-DS-R1 (Microsoft) · derived: fine tune; Phi-4-mini-flash-reasoning (Microsoft) · derived: distillation; Phi-4-mini-reasoning (Microsoft) · derived: distillation; R1 1776 (Perplexity) · derived: fine tune; Sonar Reasoning Pro (Perplexity) · derived: other
  10. DeepSeek-R1-Zero
    Based on
    DeepSeek-V3 · derived: other
  11. DeepSeek-V3-0324
    Successors
    DeepSeek-V3.1
    Based on
    DeepSeek-V3 · revision
    Siblings
    DeepSeek-V3-Base · inline variant of DeepSeek-V3
  12. DeepSeek-R1-0528
    Based on
    DeepSeek-R1 · revision
  13. DeepSeek-V3.1
    Predecessors
    DeepSeek-V3-0324
    Based on
    DeepSeek-V3 · derived: continued pretraining
    Variants and derived
    DeepSeek-V3.1-Terminus · revision
  14. DeepSeek-V3.1-Terminus
    Based on
    DeepSeek-V3.1 · revision
    Variants and derived
    DeepSeek-V3.2-Exp · derived: continued pretraining
    Siblings
    DeepSeek-V3.1-Base · inline variant of DeepSeek-V3.1
  15. DeepSeek-V3.2-Exp
    Successors
    DeepSeek-V3.2
    Based on
    DeepSeek-V3.1-Terminus · derived: continued pretraining
  16. DeepSeek-V3.2
    Predecessors
    DeepSeek-V3.2-Exp
    Variants and derived
    DeepSeek-V3.2-Speciale · variant
  17. DeepSeek-V3.2-Speciale
    Based on
    DeepSeek-V3.2 · variant
  18. DeepSeek-V4-Flash
    Predecessors
    DeepSeek-V3.2
    Variants and derived
    DeepSeek-V4-Flash-0731 · revision; DeepSeek-V4-Flash-Vision-Exp · derived: other
  19. DeepSeek-V4-Pro
    Predecessors
    DeepSeek-V3.2
    Variants and derived
    DeepSeek-V4-Pro-0813 · revision
  20. DeepSeek-V4-Flash-0731
    Based on
    DeepSeek-V4-Flash · revision
    Siblings
    DeepSeek-V4-Flash-Base · inline variant of DeepSeek-V4-Flash
  21. DeepSeek-V4-Pro-0813
    Based on
    DeepSeek-V4-Pro · revision
    Siblings
    DeepSeek-V4-Pro-Base · inline variant of DeepSeek-V4-Pro
  22. DeepSeek-V4-Flash-Vision-Exp
    Based on
    DeepSeek-V4-Flash · derived: other
  23. DeepSeek-V4.1-Flash

Families

About the organisation

  • Description DeepSeek (Hangzhou DeepSeek Artificial Intelligence Co., Ltd.) is a Chinese AI company that says it was founded in 2023. It describes itself as a research team focused on fundamental model research with an open-source approach, and says it releases model weights under the MIT License with a technical report for each model. It offers its models through a web chat, mobile apps and an API platform. Primary source[1] [2] [3] [4]

Company · https://www.deepseek.com/ · 23 models listed here

Sources

  1. [1] deepseek-ai (DeepSeek)

    Model hub page · DeepSeek

    Provenance
    Primary
    Availability
    Active
    Last checked
    Original ↗ https://huggingface.co/deepseek-ai DeepSeek's verified organisation page on Hugging Face. Its organisation card (the README of https://huggingface.co/spaces/deepseek-ai/README) says DeepSeek (深度求索) was founded in 2023 and is a Chinese company dedicated to making AGI a reality. Living page; content as seen on 2026-10-01.

  2. [2] DeepSeek Privacy Policy

    Other · DeepSeek

    Provenance
    Primary
    Availability
    Active
    Last checked
    Original ↗ https://cdn.deepseek.com/policies/en-US/deepseek-privacy-policy.html Names Hangzhou DeepSeek Artificial Intelligence Co., Ltd., with its registered address in China, as the provider and controller of the services. The page shows Last Update: Feb 10, 2026. Living page.

  3. [3] Model Mechanism and Training Methods of DeepSeek

    Documentation · DeepSeek

    Provenance
    Primary
    Availability
    Active
    Last checked
    Original ↗ https://cdn.deepseek.com/policies/en-US/model-algorithm-disclosure.html Linked from the Transparency Center (https://www.deepseek.com/en/transparency/) as Model Principles and Training Methodology. The page states no date. It refers to the company as Hangzhou DeepSeek Artificial Intelligence Co., Ltd.

  4. [4] DeepSeek | Into the Unknown

    Other · DeepSeek

    Provenance
    Primary
    Availability
    Active
    Last checked
    Original ↗ https://www.deepseek.com/en/ English homepage, linked from the Chinese homepage; links DeepSeek Web, app downloads, DeepSeek Harness, the API platform and the API docs. On 2026-10-01 www.deepseek.com failed the TLS handshake from the checking machine, while https://deepseek.com/en/ served the same page; the site declares www.deepseek.com canonical. Living page. Checked on 2026-10-05 through deepseek.com without www (same page, HTTP 200); the www host failed the TLS handshake from the checking network.