COMPANY PROFILE / AI & MACHINE LEARNING
DeepSeek
A Chinese model developer that releases large-model weights under the MIT license and documents long-context efficiency in technical reports
DeepSeek draws attention in AI and machine learning because it publishes high-performing models together with their weights, technical reports and compute kernels. In April 2026 it released DeepSeek-V4 (V4-Pro with 1.6 trillion total and 49 billion active parameters, V4-Flash with 284 billion and 13 billion, both with a one-million-token context), and on 10 September it released V4.1-Flash, which understands images, through its API and on Hugging Face. In its technical report the company says that at a one-million-token context V4-Pro needs 27% of the per-token inference compute and 10% of the KV cache of V3.2 (company claim). By contrast, revenue, headcount, funding and investors could not be confirmed from official material and are not disclosed. The U.S. National Institute of Standards and Technology (NIST) CAISI published evaluations in September 2025 and May 2026, and the Italian data protection authority ordered a restriction on processing users' data in January 2025. This profile uses only official material available by 15 October 2026 (official documentation, model cards, papers, terms, and government documents).
- DeepSeek in 30 seconds, and where it stands in AI and machine learning
- The logo and how the name is written
- Where the company sits in AI and machine learning
- Scale
- Resources committed to AI and machine learning
- What it has done so far
- What comes next
- What the company is aiming for
- Partnership map
- Technologies, products and services
- The risks this company carries
- Glossary, references and claim-to-source audit
1. DeepSeek in 30 seconds, and where it stands in AI and machine learning
Beijing DeepSeek Artificial Intelligence Co.The terms of use name both companies and their affiliates as operators; the documentation footer reads "DeepSeek, Inc."
(contact address in Haidian District, Beijing)The privacy policy says the controller has a registered address in China
V4.1-FlashV4 in April 2026; V4.1-Flash on 10 September 2026
DeepSeek has its presence by building models and publishing their weights and methods. Since R1 in January 2025 it has shown in papers, model cards and GitHub code how it trains, how it makes inference efficient, and even its compute kernels. For example, the V3 technical report states a total of 2.788 million H800 GPU hours for pre-training, context extension and post-training, which at an assumed US$2 per H800 hour comes to US$5.576 million, with a note that this "excludes the costs associated with prior research and ablation experiments" Sourced. This figure is the final training run only, at a unit price the company assumed, and is not the company's total investment. Meanwhile the company's own scale (revenue, people, capital, facilities) is not disclosed at all. What is open is the models and technology, not the business finances.
2. The logo and how the name is written
A note on naming: the company name is DeepSeek. The terms of use say the service is jointly owned and operated by Hangzhou DeepSeek Artificial Intelligence Co., Ltd., Beijing DeepSeek Artificial Intelligence Co. and their affiliates, and the privacy policy (updated 10 February 2026) names Hangzhou DeepSeek Artificial Intelligence Co., Ltd. as data controller. The documentation footer reads "DeepSeek, Inc."; the relationship between the entities (parent-subsidiary, ownership) is not stated in the official material I checked. Model names take the form "brand + generation + size", such as DeepSeek-V4-Pro, DeepSeek-V4-Flash and DeepSeek-V4.1-Flash, with dated versions (for example V4-Pro-0813). The logo image is figures/logo.svg from the official GitHub organization (deepseek-ai), which heads the model cards.
Official GitHub: https://github.com/deepseek-ai3. Where the company sits in AI and machine learning
As Fig. 1 shows, DeepSeek is a model developer. It serves models through its own API and app (chat.deepseek.com) and publishes weights on Hugging Face and elsewhere. Another feature is that it co-optimises model design and the way computers are used. The V4 technical report describes compressed attention (shrinking the long-context KV cache), overlapping expert communication with computation, low-precision (FP4 and FP8) use, and its own kernels. It is not a company that visibly owns data centers or semiconductors (not confirmed), and no cloud, infrastructure or chip business could be confirmed. See also our explainers on Transformers and large language models, reasoning models, model efficiency, and inference optimization and serving. Sourced
4. Scale
DeepSeek is unlisted and does not publish revenue, profit, headcount, capital or investors in official material. The clues to scale that could be confirmed are shown below, separated by nature Sourced.
| Item | What could be confirmed | Source |
|---|---|---|
| Revenue and profit | Not disclosed (API prices are public; revenue is not) | Reference 2 |
| Headcount | Not disclosed. The V4 technical report lists "DeepSeek-AI and 318 other authors", 319 in all (not total headcount) | Reference 4 |
| Funding and investors | Not confirmed | Not stated in official material |
| Legal entities | Operated by Hangzhou DeepSeek Artificial Intelligence Co., Ltd. and Beijing DeepSeek Artificial Intelligence Co. (terms of use, updated 20 January 2025). The data controller is the former (privacy policy, updated 10 February 2026) | Reference 11, 12 |
| Public code uptake (GitHub) | As retrieved on 15 October 2026, stars were about 105,000 for DeepSeek-V3, 92,000 for R1, 13,000 for FlashMLA, 10,000 for DeepEP and 10,000 for 3FS. An indicator of interest, not of users | Reference 10 |
| Number of released models | Since April 2026 the official Hugging Face organization has published the V4 family (Pro, Flash, their Base models, DSpark, Vision-Exp, 0731, 0813, V4.1-Flash) | Reference 17 |
"The training cost of DeepSeek's models" and "DeepSeek's size as a company" are different things. The costs in its papers (US$5.576 million for V3 and US$294,000 for R1) are the GPU hours of the final training run converted at a unit price the company assumed, and exclude prior experiments, facilities and staff. The R1 supplementary information says R1-Zero used 64x8 H800 GPUs for about 198 hours and R1 about 80 hours. The number, ownership and procurement of its hardware are not stated in official material Not yet confirmed.
5. Resources committed to AI and machine learning
For each of the four resources, the table shows what could and could not be confirmed.
| Resource | What primary information confirms | Type |
|---|---|---|
| People | Headcount is not disclosed. The V4 technical report lists 319 authors (DeepSeek-AI plus 318). The affiliations and locations of authors were not checked here. The contact address is in Haidian District, Beijing (terms of use) | Sourced / headcount not disclosed |
| Money | R&D spend, capital expenditure and funding are not disclosed. Training cost is given in the papers: V3 used 2.788 million H800 GPU hours (US$5.576 million at US$2 an hour; prior research excluded) and R1 147 thousand H800 GPU hours (US$294,000). I could not find a training cost in the V4 technical report | Sourced / totals not disclosed |
| Investments, acquisitions, JVs, divestitures | Not confirmed. The privacy policy says in general terms that user information may be disclosed in corporate transactions (mergers, asset or share sales, financing, change of control), but no specific transaction is described | Not confirmed |
| Universities, research institutes, public funds | No university collaboration or public grant is stated in the material I checked. On the technical side, the V4 report says it collaborates closely with the TileLang community. The R1 paper in Nature was peer reviewed (17 September 2025). Public funding could not be confirmed | Not confirmed / paper peer reviewed |
Next, the status of the main models. Stages and figures follow the company's own wording, and performance is the developer's stated value.
| Model | Scale | Stage | Date | Source |
|---|---|---|---|---|
| DeepSeek-V4-Pro | 1.6 trillion total / 49 billion active parameters (MoE), 1M-token context, over 32 trillion pre-training tokens. Weights in mixed FP4+FP8 | Released (API and weights) | 24 April 2026 (API); GA version V4-Pro-0813 on 13 August | Ref. 1, 3, 4 |
| DeepSeek-V4-Flash | 284 billion total / 13 billion active parameters, 1M-token context | Superseded (retired) | Released 24 April 2026; re-post-trained 31 July; replaced by V4.1-Flash on 10 September | Ref. 1, 3 |
| DeepSeek-V4.1-Flash | 552 billion backbone parameters (MoE), image and text input, 1M-token context, 45 trillion pre-training tokens. Active parameters 8 billion at prefill and 16 billion at decode | Released (API and weights) | 10 September 2026 | Ref. 1, 5 |
| DeepSeek-V3.2 / V3.2-Speciale | 685 billion total parameters (Hugging Face listing). Uses DeepSeek Sparse Attention (DSA) | Released (weights) | 1 December 2025 | Ref. 1, 6 |
| DeepSeek-R1 | A model whose reasoning is strengthened by reinforcement learning; published in Nature; weights under the MIT license | Released (weights) | 20 January 2025 (paper on 17 September) | Ref. 1, 8 |
| Infrastructure code | FlashMLA, DeepEP, DeepGEMM, 3FS (a distributed file system), DualPipe and DeepSeek Harness (an agent runtime), plus DeepGEMM-Ascend and DeepEP-Ascend for Huawei Ascend. Most under the MIT license | Released | As of 15 October 2026 | Ref. 10 |
6. What it has done so far
| When | What happened | Meaning |
|---|---|---|
| 26-27 December 2024 | Released DeepSeek-V3 through its API and posted the technical report on arXiv (2.788 million H800 GPU hours) | Detailed disclosure of training cost and method |
| 20 January 2025 | Released DeepSeek-R1 (API name "deepseek-reasoner") | A reasoning model trained with reinforcement learning, with weights released |
| 30 January 2025 | The Italian Garante (data protection authority) ordered an immediate restriction on processing users' data | The start of regulators' responses (details in chapter 11) |
| 16 April 2025 | The U.S. House Select Committee on the CCP published the report "DeepSeek Unmasked" | A congressional committee report that includes allegations (chapter 11) |
| August-September 2025 | V3.1 (21 August; thinking and non-thinking in one model), the R1 paper in Nature (17 September), V3.1-Terminus (22 September), V3.2-Exp (29 September) | The R1 method became public after peer review |
| 30 September 2025 | NIST's CAISI published an evaluation of R1, R1-0528 and V3.1 | The first evaluation by a U.S. government agency (chapter 11) |
| 1 December 2025 | V3.2 and V3.2-Speciale (the company says gold-medal level at the 2025 International Mathematical Olympiad and International Olympiad in Informatics) | Introduced the sparse attention mechanism DSA |
| 24 April 2026 | Released DeepSeek-V4 (Pro, Flash) with API and weights; technical report on arXiv on 26 April | Centered on efficiency at a 1M-token context |
| 1 May 2026 | NIST's CAISI published an evaluation of V4 Pro | Chapter 11 |
| June-July 2026 | Released speculative-decoding draft models (DSpark family); updated V4-Flash with re-post-training on 31 July | Releases on faster inference |
| 13 August 2026 | GA release of V4-Pro, three levels of thinking effort, support for the OpenAI-compatible Responses API, and peak/off-peak pricing from 16 August | Agentic use and a change to the pricing scheme |
| 21 August 2026 | Offered V4-Flash-Vision-Exp (an experimental image-understanding model) through the API | Extension to image understanding |
| 10 September 2026 | Released V4.1-Flash and lowered API prices; V4 Pro API continues after 14 September | Described by the company as the smallest model of its new architecture family |
7. What comes next
14 September 2026Billing unchanged; further notice if anything changes (Change Log)
deepseek-reasonerAnnounced for discontinuation on 24 July 2026 (April 2026); what happened afterwards was not checked here
The above is expectations within the company's own notices, not results Not yet confirmed. DeepSeek has not published plans for capital expenditure, data centers, semiconductors or funding in the official material I checked.
Revenue, profit and headcount / funding and investors / the scale and sourcing of its own compute / the training cost of V4 / timing of the next model / any DeepSeek-specific public action under U.S. export controls (none found in public BIS material; BIS is the U.S. Commerce Department's Bureau of Industry and Security) / the company's view on CAISI's evaluations. None of these is stated in the material I checked.
8. What the company is aiming for in AI and machine learning
This separates what the company says directly from what we read into it. Inference marks our own reading of primary information.
| What can be read | Primary information behind it | Type |
|---|---|---|
| It puts handling long context cheaply at the center of its technology | The V4 report is subtitled "Towards Highly Efficient Million-Token Context Intelligence". At 1M tokens V4-Pro needs 27% of the inference compute and 10% of the KV cache of V3.2. V4.1-Flash is stated to use 890 bytes of KV cache per token (about a quarter of V4-Flash) | Sourced (company claim) |
| It keeps a policy of publishing weights, methods and code | The V4, V4.1-Flash and V3.2 model cards are under the MIT license. Compute kernels (FlashMLA, DeepEP, DeepGEMM and others) are published on GitHub | Sourced |
| It is putting weight on agentic use such as coding | The V4-Pro GA announcement says agent capabilities were greatly enhanced, with Responses API support adapted for Codex, the DeepSeek Harness release and an Anthropic-format API | Sourced / Inference |
| It is extending support to hardware other than NVIDIA's (Huawei Ascend) | The V4 report says its fused expert-parallel kernel was validated on both NVIDIA GPUs and Huawei Ascend NPUs. GitHub has DeepGEMM-Ascend and DeepEP-Ascend | Sourced / Inference |
| It is moving prices down and varying them by time of day | Introduced peak/off-peak pricing (off-peak at half of peak) in August 2026 and cut prices with V4.1-Flash in September | Sourced |
The basis for the reading that it is widening beyond NVIDIA is that "-Ascend" versions sit among its public code. The V4 report says the fused kernel that overlaps expert-parallel communication and computation was "validated on both NVIDIA GPUs and Huawei Ascend NPUs", and on GitHub the Ascend versions of DeepGEMM and DeepEP were updated in September and October 2026. That is consistent with the model not depending too heavily on one type of hardware. However, the type and number of machines DeepSeek uses for training are not stated in the official material I checked (the V3 and R1 papers say H800), so which hardware V4 was trained on cannot be confirmed Not yet confirmed
9. Partnership map in AI and machine learning
The relationships in the figure are technical support, release venues, hardware use and government evaluations. For NVIDIA, the V3 and R1 papers say H800 GPUs were used, but procurement history and contracts are not in official material. Examples of cloud providers offering the models (such as deployment targets shown on Hugging Face model pages) are the providers' offerings and not statements of a partnership with DeepSeek, so they are not drawn Sourced.
10. Technologies, products and services
Model prices and performance are the developer's (company's) stated values. Benchmarks are given where name, version and setting can be confirmed, and no ranking against other companies' models is made. The company's comparison tables use rival models and settings the company chose, with no guarantee that conditions match.
| Item | Official statement (developer's stated value) | What could not be confirmed |
|---|---|---|
| API prices (checked 15 October 2026) | Per million tokens, output: V4.1-Flash (deepseek-flash) US$0.60 off-peak / US$1.20 peak; V4-Pro-0813 US$1.98 / US$3.96. Cache-miss input: US$0.15 / 0.30 and US$0.66 / 1.32. Cache-hit input: US$0.003 / 0.006 and US$0.022 / 0.044. Peak is 01:00-04:00 and 06:00-10:00 UTC on weekdays (excluding Chinese public holidays). 1M-token context, maximum output 384K. Prices may change | Effect on actual revenue; effective-cost comparison with other providers |
| Benchmarks (company-reported) | V4-Pro-Max: GPQA Diamond 90.1, LiveCodeBench 93.5, Codeforces 3206, SWE Verified 80.6, Terminal Bench 2.0 67.9, BrowseComp 83.4 (V4 model card; reasoning effort "Max"; includes a table against rival models the company chose). V4.1-Flash: GPQA Diamond 90.9, HLE 36.8, Terminal-Bench 2.1 90.6, DeepSWE v1.1 74.2 (Change Log, 10 September 2026) | Independent reproduction. NIST's CAISI found lower scores than DeepSeek's self-reported evaluations (chapter 11) |
| Architecture (company description) | V4: combined Compressed Sparse Attention (CSA) and Heavily Compressed Attention (HCA), Manifold-Constrained Hyper-Connections (mHC), the Muon optimizer and mixture of experts (MoE). V4.1-Flash: a Causal Encoder-Decoder (CED), Engram conditional memory (196 billion parameters), DSpark speculative decoding, one shared plus 384 routed experts with 6 active | Training data composition; hardware used for training |
| Safety and evaluation | The R1 Nature supplementary information includes safety evaluations (Tables 6-8, including jailbreak tests). I could not find a dedicated safety-evaluation section in the V4 and V4.1-Flash model cards | Safety evaluation of the V4 family; a framework document |
| App and API | chat.deepseek.com (app and web) and the API. Both OpenAI-compatible and Anthropic-compatible APIs, the Responses API, tool calls, JSON output and thinking modes (low / high / max) | User numbers |
The core of DeepSeek's technical reports is how memory and communication are used rather than the arithmetic itself. At long context, the KV cache that holds information on past tokens grows and presses on GPU memory capacity and bandwidth. V4 curbs this by compressing attention and says it brings the KV cache to 10% of V3.2's at a 1M-token context (company claim). V4.1-Flash shrinks it further to 890 bytes per token and uses a scheme that places the cache on SSD to avoid recomputation ("On-Disk KV Cache Storage" in the V4 report). In mixture of experts (MoE), every routing to an expert causes GPU-to-GPU communication, so a fused kernel that overlaps communication and computation hides the waiting (DeepEP), and weights are lowered to FP4 and FP8 to save capacity and bandwidth Sourced. In terms of materials and equipment, the capacity and bandwidth of HBM (high bandwidth memory), interconnects between GPUs and SSD speed directly shape model design. However, the hardware configuration DeepSeek uses is not disclosed, and how much these techniques save in actual operation cannot be told from the public figures Inference. See our explainers on model efficiency, inference optimization and serving, HBM, training clusters and interconnects, and AI accelerators.
11. The risks this company carries
The following is arranged neutrally, within what public documents of government bodies and the company's official material say. Evaluations and claims by other countries' bodies are written as those bodies' own, and no conclusion is predicted here.
| Risk | Details | Type |
|---|---|---|
| Evaluations by a U.S. government agency (CAISI) | On 30 September 2025 NIST's CAISI compared R1, R1-0528 and V3.1 on 19 benchmarks with U.S. models (GPT-5, GPT-5-mini, gpt-oss, Opus 4) and concluded they "lag behind U.S. models in performance, cost, security and adoption". It reported that agents built on R1-0528 were on average 12 times more likely to follow hijacking instructions, that R1-0528 responded to 94% of overtly malicious requests under a common jailbreak technique (8% for U.S. models), and that the models echoed four times as many inaccurate CCP-aligned narratives (by CAISI's methods). In its 1 May 2026 evaluation it called V4 Pro "the most capable PRC AI model evaluated by CAISI to date", said capabilities lag the U.S. frontier by about 8 months, that DeepSeek's self-reported scores exceed CAISI's, and that V4 was more cost efficient on 5 of 7 benchmarks against the most cost-competitive U.S. reference model. These evaluations depend on the benchmarks and settings chosen, and the company's view could not be confirmed | Sourced (agency evaluation) |
| A European data protection authority | On 30 January 2025 Italy's Garante ordered an immediate restriction on the processing of personal data of users located in Italy by Hangzhou DeepSeek and Beijing DeepSeek and opened an investigation (order document). Its reasons included an insufficient reply to its request for information, a privacy policy that did not set out the legal basis for each processing activity, data being stored in the People's Republic of China (found inconsistent with the security requirements of GDPR Article 32), and the lack of a designated EU representative. The companies had replied that they had not entered the Italian market and that the GDPR did not apply. What happened afterwards was not checked here | Sourced (authority order) |
| A U.S. House committee report | On 16 April 2025 the House Select Committee on the CCP published the report "DeepSeek Unmasked". The committee claimed that over 85% of responses are manipulated, that user data flows to China, and that DeepSeek was "reportedly" developed with over 60,000 NVIDIA chips that may have been obtained in circumvention of export controls. These are the committee's claims and its citation of reports; this article does not treat them as fact. DeepSeek's response could not be confirmed | Sourced (committee claims) |
| EU AI regulation | DeepSeek's name is not found in the list of signatories on the European Commission's page for the General-Purpose AI Code of Practice (a voluntary tool; checked 15 October 2026). Signing is not mandatory and not having signed does not itself mean a breach. Providers placing models on the EU market must show compliance with AI Act obligations by other means | Sourced / not confirmed |
| Terms and data location | The privacy policy (updated 10 February 2026) says collected personal data is stored on servers in the People's Republic of China. The terms of use are governed by the laws of mainland China, with disputes brought in the court where Hangzhou DeepSeek's registered office is located. Organizations using the service need to check fit with their own data protection and procurement policies | Sourced |
| Export controls and semiconductors | The V3 and R1 papers say H800 GPUs were used. No DeepSeek-specific public export-control measure appears in the official material I checked. U.S. export controls regulate exports of AI chips such as NVIDIA's to China and supply conditions can change (a Congressional Research Service report traces the history; reference 19) | Sourced / not confirmed |
| Limits of disclosure | Because revenue, funding, investors, facilities and staff are undisclosed, business sustainability and ownership cannot be verified from outside. Figures in papers and model cards are company claims with limited independent reproduction | Not yet confirmed |
This article summarises a company and is not legal, safety or procurement advice. Use of the models is subject to national law and your organization's policies.
12. Glossary
- Open-weight
- Publishing the trained model's weights. Distinct from full open source of training data and code.
- Mixture of experts (MoE)
- A design that activates only some "experts" per input. V4-Pro activates 49 billion of 1.6 trillion.
- KV cache
- Memory holding information on past tokens during generation; it grows with context length.
- FP4 / FP8
- 4-bit and 8-bit floating point; lowers precision to save capacity and bandwidth.
- GPU hours
- A unit where one GPU used for one hour counts as one; a gauge of training cost.
- H800
- An NVIDIA AI GPU (a version of the H100 for export to China). The V3 and R1 papers state its use.
- Ascend
- Huawei's AI NPU (neural processing unit).
- CAISI
- The Center for AI Standards and Innovation at the U.S. National Institute of Standards and Technology (NIST).
- GDPR
- The EU General Data Protection Regulation (Regulation (EU) 2016/679).
13. References
- DeepSeek Official documentation, "Change Log" (checked 15 October 2026). Release dates of V4 and V4.1-Flash, benchmarks, API changes, pricing changes, notice of retirement of legacy model names. https://api-docs.deepseek.com/updates
- DeepSeek Official documentation, "Models & Pricing" (checked 15 October 2026). Model names, context length, maximum output, peak and off-peak prices. https://api-docs.deepseek.com/quick_start/pricing
- DeepSeek Hugging Face model card, "DeepSeek-V4-Pro". Parameter counts, MIT license, evaluation results. https://huggingface.co/deepseek-ai/DeepSeek-V4-Pro
- DeepSeek-AI "DeepSeek-V4: Towards Highly Efficient Million-Token Context Intelligence" (arXiv:2606.19348, submitted 26 April 2026). Architecture, long-context efficiency, validation on NVIDIA and Ascend, author count. https://arxiv.org/abs/2606.19348
- DeepSeek Hugging Face model card, "DeepSeek-V4.1-Flash". 552 billion parameters, CED, KV cache, 45 trillion tokens. https://huggingface.co/deepseek-ai/DeepSeek-V4.1-Flash
- DeepSeek Hugging Face model card, "DeepSeek-V3.2". DSA, MIT license. https://huggingface.co/deepseek-ai/DeepSeek-V3.2
- DeepSeek-AI "DeepSeek-V3 Technical Report" (arXiv:2412.19437, submitted 27 December 2024, revised 18 February 2025). 2.788 million H800 GPU hours, a 2,048-GPU H800 cluster, cost note. https://arxiv.org/abs/2412.19437
- Guo, D. et al. "DeepSeek-R1 incentivizes reasoning in LLMs through reinforcement learning", Nature 645, 633-638 (17 September 2025). DOI 10.1038/s41586-025-09422-z. https://www.nature.com/articles/s41586-025-09422-z
- Guo, D. et al. Supplementary Information to the same paper. Section 2.4.4 and Supplementary Table 4 (training cost, 64x8 H800 GPUs) and safety evaluations. https://media.springernature.com/original/springer-static/esm/art%3A10.1038%2Fs41586-025-09422-z/MediaObjects/41586_2025_9422_MOESM1_ESM.pdf
- DeepSeek GitHub organization "deepseek-ai" (retrieved 15 October 2026). FlashMLA, DeepEP, DeepGEMM, 3FS, DeepSeek Harness, Ascend repositories and star counts. https://github.com/deepseek-ai
- DeepSeek Privacy Policy (updated 10 February 2026). Data controller and data storage location. https://cdn.deepseek.com/policies/en-US/deepseek-privacy-policy.html
- DeepSeek Terms of Use (updated 20 January 2025). Operating entities, governing law, contact address. https://cdn.deepseek.com/policies/en-US/deepseek-terms-of-use.html
- NIST CAISI "CAISI Evaluation of DeepSeek AI Models Finds Shortcomings and Risks" (30 September 2025). https://www.nist.gov/news-events/news/2025/09/caisi-evaluation-deepseek-ai-models-finds-shortcomings-and-risks
- NIST CAISI "CAISI Evaluation of DeepSeek V4 Pro" (1 May 2026). https://www.nist.gov/news-events/news/2026/05/caisi-evaluation-deepseek-v4-pro
- Garante per la protezione dei dati personali Provvedimento del 30 gennaio 2025 [10098477] (restriction on DeepSeek's processing). https://www.garanteprivacy.it/home/docweb/-/docweb-display/docweb/10098477
- U.S. House Select Committee on the CCP Press release, "Moolenaar, Krishnamoorthi Unveil Explosive Report on Chinese AI Firm DeepSeek" (report "DeepSeek Unmasked", 16 April 2025). http://selectcommitteeontheccp.house.gov/media/press-releases/moolenaar-krishnamoorthi-unveil-explosive-report-chinese-ai-firm-deepseek
- DeepSeek Official Hugging Face organization "deepseek-ai" (retrieved 15 October 2026). Used to check model release dates. https://huggingface.co/deepseek-ai
- DeepSeek Logo image in the official GitHub repository (figures/logo.svg). https://github.com/deepseek-ai/DeepSeek-V2/raw/main/figures/logo.svg
- European Commission "The General-Purpose AI Code of Practice" (list of signatories, checked 15 October 2026). https://digital-strategy.ec.europa.eu/en/policies/contents-code-gpai
- Congressional Research Service "U.S. Export Controls and China: Advanced Semiconductors" (R48642, updated 19 September 2025). The history of U.S. export controls on AI chips to China, and the fact that the H800 is a version of the H100 for export to China. https://www.congress.gov/crs_external_products/R/PDF/R48642/R48642.5.pdf
14. Claim-to-source audit
| Claim in the text | Type | Source |
|---|---|---|
| Operating entities (Hangzhou / Beijing), data controller, governing law, contact address, data storage location | Sourced | Reference 11 https://cdn.deepseek.com/policies/en-US/deepseek-privacy-policy.html Reference 12 https://cdn.deepseek.com/policies/en-US/deepseek-terms-of-use.html |
| Scale of V4-Pro and V4-Flash (1.6T / 49B, 284B / 13B), 1M tokens, over 32T tokens, FP4+FP8, MIT license, the 27% and 10% claims | Sourced | Reference 3 https://huggingface.co/deepseek-ai/DeepSeek-V4-Pro Reference 4 https://arxiv.org/abs/2606.19348 (includes company claims) |
| Scale of V4.1-Flash (552B, CED, 45T tokens, 890 bytes, 8B / 16B active), 10 September 2026, API name deepseek-flash | Sourced | Reference 1 https://api-docs.deepseek.com/updates Reference 2 https://api-docs.deepseek.com/quick_start/pricing Reference 5 https://huggingface.co/deepseek-ai/DeepSeek-V4.1-Flash |
| V4 API start (24 April 2026), V4-Flash re-post-training (31 July), V4-Pro GA and peak/off-peak pricing (13 August), Vision-Exp (21 August), V4 Pro continuing (after 14 September) | Sourced | Reference 1 https://api-docs.deepseek.com/updates |
| API prices, context length, maximum output, peak hours | Sourced | Reference 2 https://api-docs.deepseek.com/quick_start/pricing (checked 15 October 2026; may change) |
| Benchmarks (V4-Pro-Max, V4.1-Flash) and architecture descriptions | Sourced | Reference 1 https://api-docs.deepseek.com/updates Reference 3 https://huggingface.co/deepseek-ai/DeepSeek-V4-Pro Reference 5 https://huggingface.co/deepseek-ai/DeepSeek-V4.1-Flash (company's stated values; independent reproduction not confirmed) |
| V4 report with 319 authors, collaboration with TileLang, validation on NVIDIA GPUs and Huawei Ascend NPUs, on-disk KV cache | Sourced | Reference 4 https://arxiv.org/abs/2606.19348 |
| V3 training cost (2.788M H800 GPU hours, US$5.576 million, 2,048 H800 GPUs, prior research excluded) | Sourced | Reference 7 https://arxiv.org/abs/2412.19437 (unit price assumed by the company) |
| R1 training cost (147K H800 GPU hours, US$294,000, 64x8 H800 GPUs), publication in Nature (17 September 2025) | Sourced | Reference 8 https://www.nature.com/articles/s41586-025-09422-z Reference 9 https://media.springernature.com/original/springer-static/esm/art%3A10.1038%2Fs41586-025-09422-z/MediaObjects/41586_2025_9422_MOESM1_ESM.pdf |
| V3.2 and V3.2-Speciale (1 December 2025), DSA, and the statement of gold-medal level at IMO and IOI | Sourced | Reference 1 https://api-docs.deepseek.com/updates Reference 6 https://huggingface.co/deepseek-ai/DeepSeek-V3.2 (company claim) |
| GitHub repositories (FlashMLA, DeepEP, DeepGEMM, 3FS, DeepSeek Harness, Ascend versions) and star counts | Sourced | Reference 10 https://github.com/deepseek-ai |
| Contents of the CAISI evaluations (30 September 2025, 1 May 2026) | Sourced | Reference 13 https://www.nist.gov/news-events/news/2025/09/caisi-evaluation-deepseek-ai-models-finds-shortcomings-and-risks Reference 14 https://www.nist.gov/news-events/news/2026/05/caisi-evaluation-deepseek-v4-pro (agency evaluation; company's view not confirmed) |
| The Garante's restriction on DeepSeek's processing (30 January 2025) and its reasons | Sourced | Reference 15 https://www.garanteprivacy.it/home/docweb/-/docweb-display/docweb/10098477 |
| Claims of the House committee report (85%, over 60,000 chips and others) | Sourced | Reference 16 http://selectcommitteeontheccp.house.gov/media/press-releases/moolenaar-krishnamoorthi-unveil-explosive-report-chinese-ai-firm-deepseek (committee claims; not treated as fact) |
| Confirmation of model release dates (V4 family, V3.2, DSpark family and others) | Sourced | Reference 17 https://huggingface.co/deepseek-ai |
| DeepSeek not appearing in the list of signatories of the EU GPAI Code of Practice | Sourced | Reference 20 https://digital-strategy.ec.europa.eu/en/policies/contents-code-gpai (signing is not mandatory) |
| The history of U.S. export controls on AI chips to China, and the H800 being a version of the H100 for export to China | Sourced | Reference 19 https://www.congress.gov/crs_external_products/R/PDF/R48642/R48642.5.pdf (not a DeepSeek-specific measure) |
| Logo image (figures/logo.svg on the official GitHub) | Sourced | Reference 18 https://github.com/deepseek-ai/DeepSeek-V2/raw/main/figures/logo.svg |
| Revenue, profit, headcount, funding, investors, facilities, V4 training cost, timing of the next model, DeepSeek-specific export-control measures | Not yet confirmed | Not stated in the official material checked. Not estimated in this article (items not confirmed) |
| The next model, "scaling to larger models", and the handling of legacy model names after retirement | Not yet confirmed | Reference 1 https://api-docs.deepseek.com/updates Expectations within the company's notices, not results |
| The readings on long-context efficiency, open publication, agentic use, support beyond NVIDIA and the direction of pricing | Inference | Reference 1 https://api-docs.deepseek.com/updates Reference 3 https://huggingface.co/deepseek-ai/DeepSeek-V4-Pro Reference 4 https://arxiv.org/abs/2606.19348 our own interpretation |
| The explanation linking KV cache, communication and low precision to HBM, interconnects and SSD | Not yet confirmed | our own explanation (general commentary, not a company disclosure) |
Last updated: 15 October 2026 / Troy Technical
The figures in this article come from DeepSeek's official documentation, model cards, technical reports, papers and terms, and from public documents of NIST, the Italian Garante and the U.S. House Select Committee. Model names, prices, contracts and corporate structure were checked on 15 October 2026 and may change. Market-research estimates and press-report-only figures are not used. This is not an assessment of the company's financial condition or of the relative merit of its products. Passages marked "Inference" are our own reading of primary information and are not statements by the company.