DeepSeek

From Systems Analysis Wiki
Jump to navigation Jump to search

DeepSeek is a Chinese artificial intelligence company based in Hangzhou, Zhejiang, that develops large language models (LLMs), reasoning and coding models, and multimodal systems. It was founded in 2023 by Liang Wenfeng, a co-founder of the quantitative hedge fund High-Flyer. DeepSeek first drew broad attention in China when DeepSeek-V2 helped trigger an API price war in May 2024, and became internationally prominent after the releases of DeepSeek-V3 and DeepSeek-R1 in December 2024 and January 2025.[1][2]

The company publishes weights for many of its models, although licenses and the amount of released training information differ between model families. Accordingly, the models are more precisely described as "open-weight" rather than uniformly "open source": the cited releases do not provide the complete training data and reproducible end-to-end training code required by definitions such as the Open Source AI Definition.[3] As of July 11, 2026, DeepSeek's newest official mainline release is the DeepSeek-V4 Preview family: V4-Pro and V4-Flash are Mixture-of-Experts models with context windows of up to one million tokens.[4][5][6]

History

DeepSeek's founder, Liang Wenfeng, co-founded High-Flyer and became its controlling shareholder. In 2023, High-Flyer announced that it would create an independent research group devoted to artificial general intelligence (AGI); that group became DeepSeek. High-Flyer supplied early financing, computing infrastructure, and research experience accumulated through its use of machine learning in quantitative trading.[1]

DeepSeek initially operated without outside venture funding. On June 3, 2026, Reuters reported, citing people familiar with the matter, that the company was preparing to raise about 50 billion yuan (approximately US$7.4 billion) in its first external financing round, at an estimated post-investment valuation of US$52–59 billion. On June 16, Reuters relayed a report by The Information that more than 50 billion yuan had been raised at a valuation exceeding US$50 billion. According to that report, most investors contributed through a limited partnership managed by Liang rather than investing directly in DeepSeek, were subject to a five-year lock-up, and received no voting rights, preserving founder control. Reuters said that it could not independently verify the report and that DeepSeek did not immediately comment.[7][8]

On June 25, 2026, DeepSeek announced plans to at least double staffing across all departments. The recruitment notice listed positions in full-stack development, algorithms, core AI research and development, product management, and deep-learning research.[9]

The May 2024 release of DeepSeek-V2 combined a sparse Mixture-of-Experts architecture with Multi-Head Latent Attention and substantially lower API prices than many contemporary Chinese services. Reuters reported that the launch helped start a price war among Chinese model providers.[1][10]

DeepSeek attracted much wider international attention after DeepSeek-V3 and DeepSeek-R1. The V3 technical report states that the model required 2.788 million Nvidia H800 GPU-hours for pre-training, context extension, and post-training; using the paper's assumed rental rate of US$2 per GPU-hour gives a compute estimate of approximately US$5.576 million. This figure does not represent the company's total research, staffing, data, hardware, infrastructure, or earlier experimental costs, and it is not the cost of training R1.[11][12]

In a later peer-reviewed paper, DeepSeek estimated the compute cost of R1's dedicated training stages at approximately US$294,000. Its supplementary information stated that the principal R1 training run used 512 H800 GPUs for about 80 hours, after preparatory experiments with smaller models on A100 GPUs. R1 was trained from an already pre-trained DeepSeek-V3-Base checkpoint, so the US$294,000 figure does not include the much larger cost of creating that base model or broader research and infrastructure expenditure.[13][14]

By January 27, 2025, the DeepSeek assistant had overtaken ChatGPT as the most-downloaded free application in Apple's United States App Store. On the same day, investor concern that more efficient models could reduce demand for high-end AI infrastructure contributed to a global technology-stock sell-off. Nvidia shares fell just under 17%, erasing about US$593 billion in market capitalization in one trading day, while the Nasdaq Composite fell 3.1%. The market movement reflected investor expectations and uncertainty; it did not establish that DeepSeek had reproduced every capability of proprietary frontier systems at a total cost of US$6 million.[2]

Architectural Features

Mixture-of-Experts (DeepSeekMoE)
Most DeepSeek flagship language models use sparse Mixture-of-Experts layers, in which only a subset of expert subnetworks is activated for each token. DeepSeekMoE introduced fine-grained expert segmentation and shared-expert isolation. The later V3 architecture added an auxiliary-loss-free load-balancing strategy; this feature should not be retroactively attributed to every earlier DeepSeekMoE model.[15][11]
Multi-Head Latent Attention (MLA)
MLA compresses key-value representations into a lower-dimensional latent space to reduce the KV cache required during inference. In DeepSeek's V2 experiments, MLA reduced the KV cache by 93.3% relative to the company's DeepSeek 67B baseline and supported context windows of up to 128,000 tokens. The percentage is a model-specific comparison, not a universal saving for every workload.[10]
FP8 Training and Multi-Token Prediction
DeepSeek-V3 used mixed-precision FP8 training and a Multi-Token Prediction (MTP) objective, which predicts additional future tokens during training. The technical report attributes part of V3's training efficiency and downstream performance to these methods, together with extensive systems engineering.[11]
Group Relative Policy Optimization (GRPO)
DeepSeekMath introduced GRPO as a lower-memory alternative to PPO for reinforcement learning. DeepSeek-R1-Zero then applied large-scale GRPO directly to DeepSeek-V3-Base without supervised fine-tuning before the first reinforcement-learning stage, while the final R1 used a multi-stage pipeline combining cold-start data, reinforcement learning, rejection sampling, supervised fine-tuning, and a second alignment stage.[16][13]
Sparse and Compressed Attention
DeepSeek-V3.2-Exp introduced DeepSeek Sparse Attention (DSA) for more efficient long-context processing. DeepSeek-V4 uses a hybrid of Compressed Sparse Attention (CSA) and Heavily Compressed Attention (HCA), and adds Manifold-Constrained Hyper-Connections (mHC) and the Muon optimizer. At a one-million-token context length, the V4 report estimates that V4-Pro uses 27% of the single-token inference FLOPs and 10% of the KV-cache memory required by V3.2, while V4-Flash uses 10% and 7%, respectively. The report also describes FP4 quantization-aware training for routed-expert weights and the indexer QK path; other model components use higher precision.[17][5]

Model Family

  • DeepSeek LLM — bilingual 7-billion- and 67-billion-parameter base and chat models trained on 2 trillion tokens. The company's paper reported that DeepSeek LLM 67B surpassed Llama 2 70B on several evaluation sets.[18]
  • DeepSeek-Coder — code-focused models ranging from 1.3 billion to 33 billion parameters, followed by DeepSeek-Coder-V2, a Mixture-of-Experts family with 16-billion- and 236-billion-parameter variants, support for 338 programming languages, and a 128K context window.[19][20]
  • DeepSeek-V2 / V2.5 — V2 is a 236-billion-parameter MoE model with 21 billion active parameters, trained on 8.1 trillion tokens and supporting 128K context. V2.5 merged the general-chat and coding branches.[10][21]
  • DeepSeek-V3 family — the original V3 has 671 billion total and 37 billion active parameters and was pre-trained on 14.8 trillion tokens. Subsequent checkpoints included V3-0324, V3.1, V3.1-Terminus, V3.2-Exp, and V3.2, adding improved reasoning, tool use, unified thinking/non-thinking modes, and sparse attention.[11][22][23][24][25]
  • DeepSeek-R1 — a reasoning family based on DeepSeek-V3-Base. It includes R1-Zero, the final R1, the R1-0528 update, and six distilled models. R1-Zero demonstrated that verifiable reasoning behaviours could emerge from reinforcement learning without a supervised fine-tuning stage before RL; the final R1 added cold-start and alignment stages to improve readability and general capability.[13][26]
  • DeepSeek-V4 Preview — released in April 2026 in two text-model variants. V4-Pro has 1.6 trillion total parameters and 49 billion active parameters; V4-Flash has 284 billion total and 13 billion active parameters. Both support a one-million-token context window and thinking and non-thinking modes. The V4 report states that Flash and Pro were pre-trained on approximately 32 trillion and 33 trillion tokens, respectively. V4-Pro-Max is the maximum-reasoning-effort mode of V4-Pro, not a separately released checkpoint.[4][5]
  • DeepSeek-VL / VL2 and Janus — vision-language families for multimodal understanding. VL2 uses a Mixture-of-Experts architecture with variants activating approximately 1.0, 2.8, or 4.5 billion parameters. Janus and Janus-Pro unify visual understanding and image generation while using separate visual-encoding paths for the two tasks.[27][28]
  • DeepSeekMath and DeepSeek-Prover — specialized mathematical models. DeepSeekMath 7B introduced GRPO and reported 51.7% on the MATH benchmark without external tools or majority voting. DeepSeek-Prover-V2 provides 7B and 671B models for formal theorem proving in Lean 4. DeepSeekMath-V2 focuses on self-verifying mathematical proofs and verifier-guided reinforcement learning.[16][29][30]
  • DeepSeek-OCR / OCR 2 — document-understanding models that map pages into compressed visual-token sequences. OCR 2 introduced a "visual causal flow" encoder intended to reorder visual tokens according to semantic and reading relationships rather than a fixed raster scan.[31][32]
  • Distilled R1 models — six smaller checkpoints from 1.5 billion to 70 billion parameters based on Qwen 2.5 or Llama 3.1/3.3. They were fine-tuned on data generated with R1; they are not compressed copies of the 671-billion-parameter model's weights.[33]

Chronology of Key Releases

Date Release and Key Features
Nov 2, 2023 DeepSeek-Coder v1: Initial open-weight code-model family.[19]
Nov 29, 2023 DeepSeek LLM 7B/67B: Bilingual base and chat models trained on 2 trillion tokens.[18]
Jan 11, 2024 DeepSeekMoE 16B: Research model demonstrating fine-grained and shared expert specialization.[15]
Feb 5, 2024 DeepSeekMath 7B: Mathematical model and introduction of GRPO.[16]
May 6, 2024 DeepSeek-V2 236B: MoE model with 21 billion active parameters, MLA, and 128K context.[10]
Jun 17, 2024 DeepSeek-Coder-V2: 128K context and support for 338 programming languages.[20]
Sep 5, 2024 DeepSeek-V2.5: Consolidation of the general and coding model branches.[21]
Oct 17, 2024 Janus: Unified multimodal understanding and image-generation framework.[34]
Nov 20, 2024 DeepSeek-R1-Lite-Preview: Hosted preview of a reasoning model using test-time computation; DeepSeek announced that weights and API access would follow.[35]
Dec 10, 2024 DeepSeek-V2.5-1210: Final V2.5 update, improving mathematics, coding, writing, and roleplay before the transition to V3.[36]
Dec 13, 2024 DeepSeek-VL2: MoE-based vision-language family.[27]
Dec 26, 2024 DeepSeek-V3 671B: Flagship MoE model with 37 billion active parameters.[37]
Jan 20, 2025 DeepSeek-R1 / R1-Zero: Reasoning models developed with large-scale reinforcement learning.[38]
Jan 27, 2025 Janus-Pro: 1B and 7B unified multimodal understanding and image-generation models; claims of superiority over DALL-E 3 applied to selected company-reported benchmarks, not to all image-generation tasks.[28][39]
Mar 25, 2025 DeepSeek-V3-0324: Improved reasoning, front-end development, and tool-use capabilities.[22]
Apr 30, 2025 DeepSeek-Prover-V2: 7B and 671B Lean 4 theorem-proving models.[29]
May 28, 2025 DeepSeek-R1-0528: Updated reasoning checkpoint with improved function calling and structured output.[26]
Aug 21, 2025 DeepSeek-V3.1: Unified thinking and non-thinking modes and expanded agent/tool support.[23]
Sep 22, 2025 DeepSeek-V3.1-Terminus: Language-consistency and agent-use corrections.[24]
Sep 29, 2025 DeepSeek-V3.2-Exp: Experimental introduction of DeepSeek Sparse Attention.[17]
Oct 21, 2025 DeepSeek-OCR: Optical compression for document and long-context processing.[31]
Nov 27, 2025 DeepSeekMath-V2: Self-verifying mathematical reasoning and verifier-guided proof generation.[30]
Dec 1, 2025 DeepSeek-V3.2 / V3.2-Speciale: Efficient reasoning, agentic tool use, and a high-compute reasoning variant.[25]
Jan 28, 2026 DeepSeek-OCR 2: Visual Causal Flow encoder for dynamically ordered visual tokens.[32]
Apr 24, 2026 DeepSeek-V4 Preview: V4-Pro and V4-Flash with one-million-token context windows.[4]

Performance and Benchmarks

Benchmark results cited by DeepSeek are generally produced by the model developer. They depend on the exact checkpoint, prompt, sampling settings, tool access, test-time compute, and scoring metric (for example, pass@1 versus majority voting), and therefore should not be treated as independent audits or as proof of universal superiority.

  • The V3 technical report states that DeepSeek-V3 outperformed the open-weight models Llama 3.1 405B and Qwen 2.5 72B on most of its listed evaluations and was competitive with leading proprietary models on selected knowledge, mathematics, and coding benchmarks.[11]
  • The Coder-V2 paper reports that its 236-billion-parameter MoE model, after continued pre-training on an additional 6 trillion tokens, achieved results comparable to GPT-4 Turbo on several code-generation evaluations.[20]
  • DeepSeekMath 7B reported 51.7% on the MATH benchmark without external tools or majority voting. Comparisons with much larger proprietary models are benchmark-specific and do not imply equivalent general capability.[16]
  • During R1-Zero training, average AIME 2024 pass@1 increased from 15.6% to 77.9%; self-consistency over 16 samples increased the reported score to 86.7%.[13]
  • DeepSeek-Prover-V2-671B reported an 88.9% pass ratio on miniF2F-test and solved 49 of 658 PutnamBench problems.[29]
  • The Janus-Pro paper reported a GenEval score of 0.80 for Janus-Pro-7B versus 0.67 for DALL-E 3 under the authors' evaluation setup. This supports a claim about that benchmark only, not an unqualified statement that Janus-Pro "surpasses DALL-E 3" overall.[28]
  • DeepSeek's V4 report describes V4-Pro-Max, the maximum-reasoning-effort mode of V4-Pro, as the strongest open-weight system in several of its evaluations, but also reports that it remained behind leading proprietary systems on some knowledge and agent benchmarks. The same report states that V4-Pro and V4-Flash support native one-million-token contexts with substantially lower long-context compute and cache requirements than V3.2.[5]

Economics and API

DeepSeek operates a hosted API in addition to releasing model weights. As of July 11, 2026, the official API lists DeepSeek-V4-Flash at US$0.0028 per million cached input tokens, US$0.14 per million uncached input tokens, and US$0.28 per million output tokens. DeepSeek-V4-Pro is listed at US$0.003625, US$0.435, and US$0.87 per million tokens, respectively. Both provide a one-million-token context window and a maximum output length of 384,000 tokens. Prices are service terms and may change.[40]

On May 23, 2026, DeepSeek said that a 75% reduction in V4-Pro API prices would become permanent, leaving prices at one quarter of their initial level. The current dollar-denominated prices above reflect the later official pricing page rather than the approximate exchange-rate conversion reported on the announcement date.[41]

The legacy API names deepseek-chat and deepseek-reasoner currently route to the non-thinking and thinking modes of V4-Flash, respectively. DeepSeek states that those aliases are scheduled to be retired on July 24, 2026 at 15:59 UTC; clients are expected to migrate to deepseek-v4-flash or deepseek-v4-pro.[4][40]

Price comparisons with competing APIs depend on the comparison date, token type, caching rules, context length, and service tier; quoted ratios can therefore change when either provider revises its prices or product lineup.

Licensing and Open Source

DeepSeek's licensing is not uniform. The original DeepSeek-V3 repository licenses its code under MIT while placing V3 Base and Chat weights under the separate DeepSeek Model License. That model license permits commercial use, but it also contains use-based restrictions and imposes notice, attribution, and downstream redistribution obligations; it is therefore not equivalent to an unrestricted permissive software license.[42][43] The R1 repository licenses both code and the main R1 weights under MIT, although distilled checkpoints based on Qwen or Llama remain subject to the relevant upstream model terms. V3.2 and the V4-Pro and V4-Flash weight releases use the MIT license. Janus-Pro uses MIT for code but the DeepSeek Model License for model weights.[33][44][45][46][47]

The company releases model weights and selected inference code, technical reports, and evaluation materials, but the cited releases do not include the complete pre-training corpora or a fully reproducible end-to-end data and training pipeline for its frontier models. For that reason, "open-weight" is the more precise general description; the license for each checkpoint should be checked separately before use or redistribution.[42][5][3]

Industry Impact

  • DeepSeek-V2's low API price and competitive performance contributed to a price war among Chinese model providers in May 2024.[1]
  • The January 2025 popularity of V3 and R1 prompted investors to reassess assumptions about the amount of compute required for competitive models. On January 27, Nvidia lost approximately US$593 billion in market value, the largest one-day market-capitalization loss then recorded for a Wall Street company; the sell-off was a market reaction, not a technical validation of every DeepSeek claim.[2]
  • The use of Nvidia H800 accelerators for V3 and R1, and DeepSeek's later disclosure that A100 GPUs were used in preparatory R1 experiments, intensified debate over the practical effects of United States export controls. In April 2026, V4 was adapted to run on Huawei Ascend systems, and Huawei said that Ascend chips were used for part of V4-Flash training.[14][48]
  • Reuters reported in July 2026, citing unnamed sources, that DeepSeek was in the early stages of developing an inference-focused AI chip. DeepSeek had not publicly confirmed the project in the cited report, so it should not be described as a released product.[49]

Criticism and Limitations

  • Safety testing: Cisco and Robust Intelligence reported a 100% attack-success rate when an automated jailbreak method was applied to 50 randomly selected HarmBench prompts against the tested R1 configuration. This was one vendor-conducted experiment with a specific attack method and sample, not evidence that every harmful prompt succeeds in every deployment. DeepSeek's peer-reviewed R1 paper separately described the model's inherent safety as moderate and acknowledged susceptibility to jailbreaks.[50][13]
  • Security incident: In January 2025, cloud-security company Wiz found an unauthenticated DeepSeek ClickHouse database exposing more than one million log lines, including chat history, secret keys, and backend information. Wiz said DeepSeek promptly secured the exposure after responsible disclosure.[51]
  • Political filtering: The official hosted chatbot has refused, redirected, or produced formulaic answers to questions considered politically sensitive by the Chinese government, including questions about the 1989 Tiananmen Square crackdown. A 2025 United States government evaluation described by Reuters found that DeepSeek-R1 frequently used language aligned with Beijing's official positions on sensitive topics. These observations concern tested hosted or post-trained systems and should be distinguished from locally deployed open-weight checkpoints, whose outputs can differ according to the checkpoint, system prompt, fine-tuning, and moderation layer.[52]
  • Privacy and regulation: DeepSeek's privacy policy states that the hosted service directly collects, processes, and stores personal data in the People's Republic of China and may use interactions to improve and train its technology. These terms concern the official cloud service and do not automatically apply when a downloaded model is run entirely on infrastructure controlled by the user. Italy's data-protection authority ordered an immediate limitation on the processing of Italian users' data in January 2025; Australia directed government entities to remove DeepSeek products from government systems; South Korea's privacy authority later said that user information and prompts had been transferred without the required consent during the service's initial launch; and several United States Commerce Department bureaus prohibited access on government-furnished devices. A separate Italian consumer-protection investigation was closed in January 2026 after DeepSeek made binding commitments to warn users more clearly about hallucination risks; that closure did not reverse the earlier privacy authority's order.[53][54][55][56][57][58]
  • Distillation and intellectual-property allegations: In April 2026, Reuters reported that the United States State Department had instructed diplomatic posts to raise allegations that DeepSeek and other Chinese companies had used unauthorized model distillation or extracted intellectual property from United States AI laboratories. OpenAI had made related allegations. The Chinese Embassy rejected the accusations, and DeepSeek had previously said that V3 used naturally occurring web-crawled data and that it had not intentionally used OpenAI-generated synthetic data. The cited report described government and company allegations, not a judicial finding of liability.[59]
  • Model limitations: The R1 paper reported excessive reasoning on simple questions, language mixing outside Chinese and English, sensitivity to prompting, and initially weak structured output and tool use. R1-0528 and later V3/V4 models improved several of these areas, but the limitations remain relevant to the original R1 checkpoint. The V4 report also notes that the preview series can make mistakes on simple tasks, misinterpret vague instructions, overthink some questions, and trail some proprietary frontier models on selected knowledge and agent evaluations.[13][26][5]

See also

Literature

  • Bi, X. et al. (2024). DeepSeek LLM: Scaling Open-Source Language Models with Longtermism. arXiv:2401.02954.
  • Dai, D. et al. (2024). DeepSeekMoE: Towards Ultimate Expert Specialization in Mixture-of-Experts Language Models. arXiv:2401.06066.
  • Shao, Z. et al. (2024). DeepSeekMath: Pushing the Limits of Mathematical Reasoning in Open Language Models. arXiv:2402.03300.
  • DeepSeek-AI (2024). DeepSeek-V2: A Strong, Economical, and Efficient Mixture-of-Experts Language Model. arXiv:2405.04434.
  • Zhu, Q. et al. (2024). DeepSeek-Coder-V2: Breaking the Barrier of Closed-Source Models in Code Intelligence. arXiv:2406.11931.
  • Wu, Z. et al. (2024). DeepSeek-VL2: Mixture-of-Experts Vision-Language Models for Advanced Multimodal Understanding. arXiv:2412.10302.
  • DeepSeek-AI (2024). DeepSeek-V3 Technical Report. arXiv:2412.19437.
  • Chen, X. et al. (2025). Janus-Pro: Unified Multimodal Understanding and Generation with Data and Model Scaling. arXiv:2501.17811.
  • Guo, D. et al. (2025). DeepSeek-R1 incentivizes reasoning in LLMs through reinforcement learning. Nature.
  • Ren, Z. Z. et al. (2025). DeepSeek-Prover-V2: Advancing Formal Mathematical Reasoning via Reinforcement Learning for Subgoal Decomposition. arXiv:2504.21801.
  • Wei, H. et al. (2025). DeepSeek-OCR: Contexts Optical Compression. arXiv:2510.18234.
  • Shao, Z. et al. (2025). DeepSeekMath-V2: Towards Self-Verifiable Mathematical Reasoning. arXiv:2511.22570.
  • Wei, H. et al. (2026). DeepSeek-OCR 2: Visual Causal Flow. arXiv:2601.20552.
  • DeepSeek-AI (2026). DeepSeek-V4: Towards Highly Efficient Million-Token Context Intelligence. arXiv:2606.19348.

References

  1. ↑ 1.0 1.1 1.2 1.3 High-Flyer, the AI quant fund behind China's DeepSeek // Reuters. 2025-01-29.
  2. ↑ 2.0 2.1 2.2 DeepSeek sparks AI stock selloff; Nvidia posts record market-cap loss // Reuters. 2025-01-27.
  3. ↑ 3.0 3.1 The Open Source AI Definition — 1.0 // Open Source Initiative.
  4. ↑ 4.0 4.1 4.2 4.3 DeepSeek V4 Preview Release // DeepSeek API Docs. 2026-04-24.
  5. ↑ 5.0 5.1 5.2 5.3 5.4 5.5 DeepSeek-V4: Towards Highly Efficient Million-Token Context Intelligence // arXiv. 2026.
  6. ↑ Change Log // DeepSeek API Docs. Accessed 2026-07-11.
  7. ↑ DeepSeek slated to raise $7 billion in maiden funding round, sources say // Reuters. 2026-06-03.
  8. ↑ China's DeepSeek closes over $7 billion funding with unusual deal structure, the Information reports // Reuters. 2026-06-16.
  9. ↑ China's DeepSeek plans to at least double staff in all departments // Reuters. 2026-06-25.
  10. ↑ 10.0 10.1 10.2 10.3 DeepSeek-V2: A Strong, Economical, and Efficient Mixture-of-Experts Language Model // arXiv. 2024.
  11. ↑ 11.0 11.1 11.2 11.3 11.4 DeepSeek-V3 Technical Report // arXiv. 2024.
  12. ↑ DeepSeek's low-cost AI spotlights billions spent by US tech // Reuters. 2025-01-27.
  13. ↑ 13.0 13.1 13.2 13.3 13.4 13.5 DeepSeek-R1 incentivizes reasoning in LLMs through reinforcement learning // Nature. 2025.
  14. ↑ 14.0 14.1 China's DeepSeek says its hit AI model cost just $294,000 to train // Reuters. 2025-09-18.
  15. ↑ 15.0 15.1 DeepSeekMoE: Towards Ultimate Expert Specialization in Mixture-of-Experts Language Models // arXiv. 2024.
  16. ↑ 16.0 16.1 16.2 16.3 DeepSeekMath: Pushing the Limits of Mathematical Reasoning in Open Language Models // arXiv. 2024.
  17. ↑ 17.0 17.1 DeepSeek-V3.2-Exp Release // DeepSeek API Docs. 2025-09-29.
  18. ↑ 18.0 18.1 DeepSeek LLM: Scaling Open-Source Language Models with Longtermism // arXiv. 2024.
  19. ↑ 19.0 19.1 DeepSeek-Coder: When the Large Language Model Meets Programming — The Rise of Code Intelligence // arXiv. 2024.
  20. ↑ 20.0 20.1 20.2 DeepSeek-Coder-V2: Breaking the Barrier of Closed-Source Models in Code Intelligence // arXiv. 2024.
  21. ↑ 21.0 21.1 DeepSeek-V2.5 Release // DeepSeek API Docs. 2024-09-05.
  22. ↑ 22.0 22.1 DeepSeek-V3-0324 Release // DeepSeek API Docs. 2025-03-25.
  23. ↑ 23.0 23.1 DeepSeek-V3.1 Release // DeepSeek API Docs. 2025-08-21.
  24. ↑ 24.0 24.1 DeepSeek V3.1 Update // DeepSeek API Docs. 2025-09-22.
  25. ↑ 25.0 25.1 DeepSeek-V3.2 Release // DeepSeek API Docs. 2025-12-01.
  26. ↑ 26.0 26.1 26.2 DeepSeek-R1-0528 Release // DeepSeek API Docs. 2025-05-28.
  27. ↑ 27.0 27.1 DeepSeek-VL2: Mixture-of-Experts Vision-Language Models for Advanced Multimodal Understanding // arXiv. 2024.
  28. ↑ 28.0 28.1 28.2 Janus-Pro: Unified Multimodal Understanding and Generation with Data and Model Scaling // arXiv. 2025.
  29. ↑ 29.0 29.1 29.2 DeepSeek-Prover-V2: Advancing Formal Mathematical Reasoning via Reinforcement Learning for Subgoal Decomposition // arXiv. 2025.
  30. ↑ 30.0 30.1 DeepSeekMath-V2: Towards Self-Verifiable Mathematical Reasoning // arXiv. 2025.
  31. ↑ 31.0 31.1 DeepSeek-OCR: Contexts Optical Compression // arXiv. 2025.
  32. ↑ 32.0 32.1 DeepSeek-OCR 2: Visual Causal Flow // arXiv. 2026.
  33. ↑ 33.0 33.1 DeepSeek-R1 // GitHub.
  34. ↑ Janus: Decoupling Visual Encoding for Unified Multimodal Understanding and Generation // arXiv. 2024.
  35. ↑ DeepSeek-R1-Lite-Preview Release // DeepSeek API Docs. 2024-11-20.
  36. ↑ DeepSeek-V2.5-1210 Release // DeepSeek API Docs. 2024-12-10.
  37. ↑ Introducing DeepSeek-V3 // DeepSeek API Docs. 2024-12-26.
  38. ↑ DeepSeek-R1 Release // DeepSeek API Docs. 2025-01-20.
  39. ↑ DeepSeek's Janus-Pro AI model beats rivals in image generation // Reuters. 2025-01-27.
  40. ↑ 40.0 40.1 Models & Pricing // DeepSeek API Docs. Accessed 2026-07-11.
  41. ↑ China's DeepSeek to make permanent 75% price cut on flagship V4-Pro AI model // Reuters. 2026-05-23.
  42. ↑ 42.0 42.1 DeepSeek-V3 // GitHub.
  43. ↑ DeepSeek Model License // GitHub.
  44. ↑ DeepSeek-V3.2 // Hugging Face.
  45. ↑ DeepSeek-V4-Pro License // Hugging Face.
  46. ↑ DeepSeek-V4-Flash License // Hugging Face.
  47. ↑ Janus-Pro-7B // Hugging Face.
  48. ↑ DeepSeek-V4, the Chinese AI model adapted for Huawei chips // Reuters. 2026-04-24.
  49. ↑ China's DeepSeek developing its own AI chip, sources say // Reuters. 2026-07-07.
  50. ↑ Evaluating Security Risk in DeepSeek and Other Frontier Reasoning Models // Cisco. 2025-01-31.
  51. ↑ Wiz Research Uncovers Exposed DeepSeek Database Leaking Sensitive Information, Including Chat History // Wiz. 2025-01-29.
  52. ↑ US scrutinizes Chinese AI for ideological bias, memo shows // Reuters. 2025-07-09.
  53. ↑ DeepSeek Privacy Policy // DeepSeek. Accessed 2026-07-11.
  54. ↑ Artificial Intelligence: The Italian Data Protection Authority blocks DeepSeek // Garante per la protezione dei dati personali. 2025-01-30.
  55. ↑ Direction 001-2025 on DeepSeek Products, Applications and Web Services // Australian Government. 2025-02-04.
  56. ↑ South Korea says DeepSeek transferred user data, prompts without consent // Reuters. 2025-04-24.
  57. ↑ US Commerce Department bureaus ban China's DeepSeek on government devices, sources say // Reuters. 2025-03-17.
  58. ↑ Italy closes probe into DeepSeek after commitments to warn of AI hallucination risks // Reuters. 2026-01-05.
  59. ↑ US State Department orders global warning about alleged AI thefts by DeepSeek, other Chinese firms // Reuters. 2026-04-24.
Retrieved from "https://systems-analysis.ru/eng/index.php?title=DeepSeek&oldid=3130"