DeepSeek

What DeepSeek is and what the name covers

DeepSeek refers to an artificial intelligence research company, its model families and a conversational service. The application works with instructions and materials; the API integrates models into other programs; published weights enable independent deployments. These are different access routes: a repository feature is not necessarily available in every conversation in the app.

Its models generate responses from context by predicting tokens. Pretraining learns patterns and subsequent tuning shapes instruction following. This profile separates technical history, capabilities documented on 8 November 2026 and occupational safety and health (OSH) uses proposed by Sabentis.

Reference documentation: Identity and model training disclosure · DeepSeek API models and pricing.

Origins, leadership and financing

Liang Wenfeng established DeepSeek in 2023 within the orbit of quantitative fund High-Flyer in Hangzhou. The research drew on earlier experience in machine learning and computing infrastructure. Reuters documented that background: this was not a project that emerged without previous investment when R1 became popular.

In June 2026, TechRadar reported a first external funding round of about $7.4 billion, valuing the company above $50 billion, and a recruitment campaign. These are attributed press figures, not accounts audited for this profile. The stated ambition to develop artificial general intelligence does not establish that it has been achieved.

Reference documentation: Reuters: High-Flyer, the AI quant fund behind China's DeepSeek · TechRadar: DeepSeek hiring and financing, June 2026.

From Coder and LLM to Math and V2

• Coder, 2023: programming models in different sizes, trained primarily on code. The release distinguishes the code licence from the model licence.
• DeepSeek-LLM: the 7-billion- and 67-billion-parameter Base and Chat variants organised the general language offering, with a particular emphasis on English and Chinese.
• DeepSeekMoE and Math: the first line investigated shared and selected experts; the second started from Coder to train mathematical capabilities through several stages, including reinforcement learning.
• V2, May 2024: combined mixture of experts and Multi-head Latent Attention, with 236 billion total parameters and 21 billion active per token. V2.5 continued bringing conversation and programming together.

This progression explains why “DeepSeek” does not identify one unchanging model. Training, architecture and access methods differ across generations.

Reference documentation: DeepSeek-Coder repository · DeepSeek-LLM repository · DeepSeekMoE repository · DeepSeekMath repository · DeepSeek-V2 repository · DeepSeek API change log.

Mixture of experts, attention and V3

A mixture of experts, or MoE, distributes some computation among specialised network blocks. Routing selects only some of them for each token. “Expert” is a mathematical term: it does not represent a person or establish professional competence in prevention.

V2’s latent attention compresses information used to maintain context. V3, released in December 2024, increased the scale to 671 billion parameters, with 37 billion active per token, and introduced changes to load balancing and multiple-token prediction. Total size, active computation and required memory are different measures; none should become an automatic quality ranking.

Reference documentation: DeepSeekMoE repository · DeepSeek-V2 repository · DeepSeek-V3 repository.

Infrastructure and the cost debate

The V3 report estimated $5.576 million for the training described, assuming H800 GPU rental at two dollars per hour. That figure is not the entire cost of creating DeepSeek: it does not encompass all preceding research, facilities, staffing and experiments.

Fire-Flyer AI-HPC research describes hardware and software co-design, with HFReduce for communication, HaiScale for distributed training, 3FS for files and HAI-Platform for management. Efficiency involves coordination and data movement as well as the model.

For an organisation, useful cost is measured per accepted piece of work: preparation, inference, review, corrections, hosting and maintenance. A low token price can become less important when much of the output needs to be rewritten.

Reference documentation: DeepSeek-V3 Technical Report · Fire-Flyer AI-HPC research.

R1 and developments during 2025

R1-Zero investigated reinforcement learning without initial supervised tuning, encountering readability problems. R1 added preparation and tuning; its Qwen- and Llama-based distillations are different variants, not identical copies of the large model.

R1 launched on 20 January 2025 and was updated in May. V3.1 arrived in August, Terminus in September and V3.2 in December. V3.2 developed sparse attention, reinforcement training and synthetic tasks for tool use; Speciale prioritised reasoning with more computation.

Competition and benchmark results published by the manufacturer describe those protocols. They do not establish that an answer about a particular workplace is correct.

Reference documentation: DeepSeek-R1 repository · DeepSeek API change log · DeepSeek-V3.2 technical report.

V4 and V4.1: architectural changes in 2026

V4 appeared in preview in April 2026, with Pro and Flash and a one-million-token context. Its design combined attention compression mechanisms, mHC and Muon optimisation.

V4.1-Flash introduced a causal encoder-decoder architecture and native visual input in September. Its CSA2 attention selects information from context and Engram incorporates token-table retrieval. This architectural “memory” is not a notebook of personal memories about the user. The Flash version checked here corresponds to the 17 October update.

A long window increases the material that can be submitted; it does not guarantee that every exception hidden within that material will be identified.

Reference documentation: DeepSeek-V4-Pro model card · DeepSeek-V4.1-Flash model card · DeepSeek API change log.

Application, API and independent deployment

The API maps deepseek-flash to V4.1-Flash-1017 and deepseek-v4-pro to V4-Pro-0813. Only Flash supports vision. Pro remains available despite the earlier retirement announcement.

Billing distinguishes input, cached and output tokens, plus Search/Fetch requests. Before budgeting, consult the applicable rates and measure the actual workflow.

R1 and V4 were released under the MIT licence. Check the particular model, its components and the terms governing derived variants. Downloading weights does not itself provide servers, support, logs, access control or a finished enterprise service.

Reference documentation: DeepSeek API models and pricing · DeepSeek-R1 repository · DeepSeek-V4-Pro model card · DeepSeek API change log.

What thinking mode adds

The API allows thinking mode to be enabled or disabled and its effort adjusted. It is enabled by default. In conversations involving tools, the documentation requires reasoning content to be retained and resubmitted according to the protocol; ignoring this requirement can break an integration.

For editorial work, ask for detected conditions and contradictions before requesting a proposal. A detailed explanation can help inspect the result, but it is not independent proof. The test is whether the answer preserves the data, identifies unknowns and allows checks to be reproduced.

Reference documentation: DeepSeek thinking mode.

Images, documents and multimodal families

Flash accepts images alongside text and can describe them or read charts. The API supports URLs, encoded data and uploaded files, with limits and resizing. A small photograph may lose precisely the detail needed to interpret a sign.

DeepSeek-VL2 is a dedicated visual understanding, OCR and document model line. Janus, JanusFlow and Janus-Pro investigate image understanding and generation; Janus-Pro’s January 2025 release does not mean all those functions are integrated into today’s conversational application.

In OSH, an image can help prepare discussion questions. An incomplete visual observation should not become certification of a machine’s condition or a person’s exposure.

Reference documentation: DeepSeek vision guide · Janus repository · DeepSeek-VL2 repository.

Search, sources and verifiable answers

Search returns structured web results; Fetch retrieves a page in Markdown. These services support search workflows and systems that retrieve documents before answering. Searching, retrieving and generating are separate operations: a successful search can still lead to a misinterpreted conclusion.

For prevention research, the proposed workflow retains the title, link, date and passage supporting each assertion. Review the original page and separate its statements from interpretation. A nonexistent reference, an obsolete rule or guidance for a different sector must be excluded, even when the generated wording sounds persuasive.

Reference documentation: DeepSeek Search and Fetch.

Code, formal proof and agents

DeepSeek-Prover-V2 works on formal theorem proving in Lean 4. This field requires proofs that a formal system can verify and differs from mathematical explanation in everyday language.

DeepSeek Harness is an agent environment with a Cordis plugin architecture and MIT licensing. Its interface allows a workspace to be selected; the agent can read and edit files, run commands and delegate tasks. Confirmations depend on the active permission policy.

For OSH prototypes, use an isolated repository and fictional data. A function that calculates correctly for three examples may fail with empty values, different units or duplicate records. Acceptance should include those cases before any operational data is connected.

Reference documentation: DeepSeek-Prover-V2 repository · DeepSeek Harness repository · DeepSeek Harness: Web UI guide.

Privacy, location and usage rights

The February 2026 policy places data processing and storage in China and states that the service is not designed to receive sensitive personal data, including health information. It also provides controls for opting out of data use for model improvement.

The March 2026 terms distinguish input and output rights, warn that answers may not be unique and that shared links allow access to recipients. A third-party application may additionally apply its own terms.

In prevention work, removing a name is insufficient if the company, role, shift and remaining details identify someone. The proposed workflow uses invented situations or authorised, minimised materials. Independent hosting changes who manages data, but requires checks on logs, backups, extensions and outbound connections.

Reference documentation: DeepSeek privacy policy · DeepSeek terms of use.

Security incident and data protection action

In January 2025, Wiz reported an unauthenticated ClickHouse database exposing DeepSeek logs, conversations and keys. According to its report, the exposure was closed after the company was notified. The incident documents that problem; it does not prove every current installation remains affected.

South Korea’s PIPC published recommendations in April 2025 covering international transfers, data sent to Volcano and transparency. Separately, Italy’s Garante ordered restrictions on processing data of people in Italy on 30 January 2025.

These were different actions involving distinct authorities, dates and subjects. They should not be reduced to a single worldwide ban or treated as later approval of every use.

Reference documentation: Wiz: exposed DeepSeek database · PIPC: DeepSeek examination results · Garante: decision of 30 January 2025.

Institutional restrictions: scope matters

The US notice of 20 October 2026 excludes DeepSeek products from federal civilian agency procurement and systems, including third-party implementations. Its scope does not amount to a prohibition on every individual’s use of the product.

Victoria’s February 2025 direction, following Australian federal action, establishes government restrictions and specific conditions for inspectable local models. One institution’s details cannot be extrapolated to another.

AGCM closed its reliability-disclosure investigation with commitments, without finding an infringement. This action is separate from the Garante’s.

A purchasing decision must start with the country, sector, contract and policy applicable to the actual deployment. This profile describes documented decisions; it does not determine each organisation’s authorisation.

Reference documentation: US Federal Register: DeepSeek removal order notice · Victoria: direction on DeepSeek · AGCM: commitments on hallucination disclosures.

Public reception and international competition

Attention to R1 extended beyond technology. On 27 January 2025, it coincided with a technology-stock sell-off: Nvidia fell almost 17%. VOA reported opposing views on efficiency and US chip-export controls. The episode placed DeepSeek within debate about technological competition between China and the United States.

A stock-market reaction expresses investor expectations, not a measurement of accuracy. Similarly, a downloadable model allows examination of aspects that a closed service does not expose, but does not necessarily reveal its entire training corpus. Openness, performance and governance should be assessed separately.

Reference documentation: VOA: China's DeepSeek AI rattles Wall Street, but questions remain.

Censorship, distillation and critical reception

Qiu, Zhou and Ferrara analysed 646 politically sensitive prompts and documented omissions and reformulations between reasoning and answers. The study supports examining moderation and bias in the evaluated models; it does not automatically establish identical behaviour in every later release.

In September 2026, Anthropic attributed a reasoning-extraction campaign and redirection of selected requests to Claude to DeepSeek. These are the provider’s allegations based on its investigation, not a judicial finding or a Sabentis verification.

International reception combined interest in efficiency and downloadable models with questions about costs, controls and data provenance. A useful comparison identifies the version, hosting arrangement, date and task. Market headlines do not replace an evaluation of quality in the work actually intended.

Reference documentation: Qiu, Zhou and Ferrara: Information Suppression in Large Language Models · Anthropic: September 2026 threat intelligence report.

Relationship to occupational safety and health

The following uses are editorial proposals from Sabentis, not demonstrated manufacturer outcomes. DeepSeek can support linguistic or technical preparation where a controlled source exists and someone can review the result.

• Training: turn an authorised procedure into questions, situations and justified answers. The intended benefit is having different ways to explain the same rule while preserving its meaning.
• Document preparation: identify ambiguities between two supplied instructions. Request discrepancies with the original passage, without guessing which instruction should prevail.
• Analytical support: draft a formula or small script using a fictional table. Check denominators, periods, units and empty cases before reusing it.

Decisions about fitness for work, diagnosis, work authorisation or critical preventive measures are not delegated. The tool’s value lies in preparing inspectable material, not in giving authority to generated text.

Reference documentation: Identity and model training disclosure · DeepSeek vision guide · DeepSeek Search and Fetch.

Hypothetical case: preparing a housekeeping session

A team wants to prepare a ten-minute discussion about a storage area. It starts from an invented description: packaging remains beside a walkway; at the end of the shift, two people interpret responsibility for removing it differently. A fictional procedure with three rules and a specified responsible role is also supplied.

The input avoids names, identifiable photographs, plans and accident data. The assignment requests six open questions and three decision scenarios, indicating the rule used for each. A specific editorial instruction is added: where information is missing, ask a question; do not fill out the procedure with assumed obligations.

The expected result is a discussion draft, not a new work instruction. That distinction allows it to be checked without accidentally introducing a rule that nobody has approved.

Reference documentation: Identity and model training disclosure.

Developing and reviewing the exercise

1. Extract rules: ask for the three supplied rules to be listed and reproduced as brief references without expansion. Compare them with the document before generating activities.
2. Create situations: request one scene with enough information, another with a missing detail and another with a contradiction. This tests whether uncertainty is recognised instead of always producing a definitive answer.
3. Review proposals: check that no answer blames an individual, invents a penalty or introduces equipment and substances absent from the case.
4. Adapt language: prepare Spanish, Catalan and English using the same glossary. A bilingual reviewer should verify meaning before the translations are used with workers.
5. Prepare the session: choose three useful questions, save the reviewed version and remove any question whose answer depends on a nonexistent rule.

If the model proposes “always remove immediately” when the instruction requires notifying a responsible person, the draft fails even if the sentence sounds cautious. Review tests fidelity to the case, not just grammar.

Reference documentation: DeepSeek thinking mode.

What would establish that the material is useful?

The proposed acceptance criteria require every answer to link to a supplied rule, missing information to remain flagged and all three translations to preserve responsibilities and conditions. Record what was corrected and why. Do not assign improvement percentages without measuring them.

In a later trial, someone who did not write the material could explain what they would do in each scene and locate the corresponding rule. If they must guess missing information, revise the exercise. If they understand the text but the real situation works differently, first correct the source with the competent responsible person.

This case has not been carried out or validated with workers. Its purpose is to provide a reproducible procedure for assessing a draft before making it training material.

When to choose it and what to combine it with

DeepSeek is worth evaluating when its API-accessible models, published research or a maintainable independent deployment match the task. For occasional work, consider whether preparing an integration takes more effort than manually reviewing a short text.

A comparison with ChatGPT can use the same instruction, sources and acceptance criteria, without declaring a winner from one answer. For a document collection, the NotebookLM profile helps examine another approach. Reproducible calculations can remain in Excel, leaving explanation or formula drafting to AI.

Selection should document quality, review time, full cost, data handling and the ability to leave the tool while retaining useful materials.

Reference documentation: DeepSeek Harness: Web UI guide.

Frequently asked questions

Is R1 still the name of the current service?

R1 is an important historical family. For a current integration, check the identifier and served version; retaining an old name in a program does not guarantee use of that model.

Does a downloaded model automatically provide privacy?

No. Assessment must cover the environment running it, its logs, connections and users. Local control works only when that entire chain is configured and maintained.

Can it produce a complete risk assessment?

It can help structure questions and drafts using authorised information. Observing work, selecting methods and validating conclusions require contextual knowledge and professional responsibility. Producing a long document does not establish that competence.

Reference documentation: DeepSeek API models and pricing · DeepSeek-R1 repository · DeepSeek privacy policy.