DeepSeek's V4 training disclosure changes the open-weights conversation.

What the 64 pages actually contain

The document is organised into eleven sections. Section one covers the data composition: 14.8 trillion tokens of pre-training data, split across 71 per cent web text, 14 per cent code, 8 per cent mathematics and scientific literature, 4 per cent multilingual text covering 38 non-English languages, and 3 per cent synthetic data generated by DeepSeek's own V3 model. The web text component is decomposed further: 41 per cent Common Crawl 2024 snapshots, 22 per cent the ZGLM-Web-2 corpus (a Chinese-language web index that DeepSeek maintains internally), 18 per cent academic and reference materials cleared for redistribution, 12 per cent licensed text from named publishers including Springer Nature, IEEE, and a consortium of Chinese state press operations, and 7 per cent from sources the document describes as "domain-specific public datasets" with a per-source enumeration that runs across 91 named corpora. The per-source enumeration is the structural shift. No previous frontier disclosure has included it at this granularity.

Section two covers tokeniser construction. V4 uses a 256,000-token vocabulary built on a byte-pair encoding extension that DeepSeek calls Byte-Aware BPE-2. The tokeniser was trained on a 3.2-trillion-token subsample drawn proportionally from the pre-training corpus, with explicit balancing rules to prevent overrepresentation of Chinese-language morphology that the team identified as a known failure mode in the V3 tokeniser. The vocabulary distribution is published: 41 per cent English, 19 per cent Chinese, 14 per cent code-syntactic tokens, 11 per cent mathematical notation, 8 per cent multilingual (38 languages), and 7 per cent control and structural tokens. The tokeniser specification is precise enough that an independent team can reproduce the vocabulary from the published methodology — a property that no previous OpenAI, Anthropic, or Google frontier release has supplied at comparable detail.

Sections three through five cover the pre-training architecture, the optimiser configuration, and the compute allocation. V4 uses a mixture-of-experts architecture with 1.42 trillion total parameters and 81 billion active parameters per forward pass, distributed across 192 routed experts and 8 shared experts. The MoE routing follows what the document calls "balanced auxiliary-loss-free routing" — a refinement of the V3 routing protocol that the team claims reduces expert imbalance by 38 per cent on the published benchmark suite. The optimiser is Adam-variant with decoupled weight decay and a learning rate schedule that is plotted across the full training run. The compute allocation is the section that has drawn the most attention: 4.92 million H800 GPU-hours for pre-training, 1.18 million H800 GPU-hours for post-training, and 312,000 H800 GPU-hours for evaluator-cohort experiments and ablations. The total at 6.41 million H800 GPU-hours sits roughly 19 per cent below what independent analysts at SemiAnalysis and Epoch AI had estimated DeepSeek would require for a V4-scale run. The compute figure is the figure that procurement teams in US Fortune 500 buyers triangulated against the disclosed parameter count to assess credibility. The triangulation, by every published analyst response within 36 hours, lands inside the plausibility band.

Sections six through nine cover the post-training pipeline. RLHF protocol design occupies 14 pages of the document, more than any other section. The disclosure includes the evaluator panel composition (94 internal evaluators, 612 contracted evaluators across 17 countries, 28 per cent of contracted evaluators native Mandarin, 41 per cent native English, 31 per cent native other languages), the preference data collection protocol (a paired-comparison format with named rubrics for helpfulness, harmlessness, accuracy, and what DeepSeek calls "directional reasoning fidelity"), the reward model architecture (a 32-billion-parameter dense transformer trained on 4.1 million paired preferences), the PPO configuration, and the constitutional rules library that DeepSeek applied during the late-stage refinement. The constitutional library is published as an appendix and runs to 188 named principles in English and Chinese. The publication of the constitutional library is the section that legal teams inside Western buyers have spent the most time reviewing, because it represents the first public enumeration of the value-shaping principles a Chinese-origin frontier model has been trained against.

Sections ten and eleven cover the evaluation methodology and the safety analysis. The evaluation suite is published with full per-benchmark numbers across MMLU, GPQA, MATH, HumanEval, SWE-Bench Verified, GAIA, AgentBench, and a Chinese-language evaluation suite that DeepSeek introduces in the document. The safety analysis runs through dual-use risk assessment, biosecurity evaluation, cybersecurity capability evaluation, and what the document calls "long-horizon agency risk" — a category that maps closely to the Anthropic and Google DeepMind frameworks for assessing models that operate in extended agent loops. The safety section is the section that two of the Western reproducibility teams have flagged as the most consequential disclosure surface, because it provides a structured comparison point against the safety evaluations that frontier Western labs have published in their respective model cards.

The procurement reaction at Fortune 500 buyers

Within 72 hours of the disclosure, at least 18 Fortune 500 procurement desks had circulated internal memos addressing DeepSeek V4 evaluation posture. The memos cluster into three patterns. The first pattern, present in roughly 9 of the 18 memos reviewed, treats the disclosure as a procurement-neutral event: V4 is open-weights, the methodology is documented at unprecedented detail, and the buyer's existing prohibition on Chinese-origin model deployment remains in force because the prohibition is jurisdictional rather than methodological. Bank of America's procurement memo, which was distributed on 19 May 2026 across the bank's technology and risk functions, falls into this pattern. The memo states that "the V4 disclosure is technically credible and does not modify our position. The position is set by the data-egress and jurisdictional review framework, which the disclosure does not address." The procurement teams writing these memos read the disclosure carefully, acknowledged the technical credibility, and did not modify their stance.

The second pattern, present in approximately 6 of the 18 memos, treats the disclosure as procurement-relevant but procurement-deferred. The memos in this pattern direct the technology and risk teams to convene a working group to evaluate whether the disclosure affects internal deployment guidelines, but explicitly state that no production deployment decision is implied. ExxonMobil's IT risk office issued a memo of this kind on 20 May 2026, directing a 90-day working group review with no commitment to a production change at the conclusion. JPMorgan Chase's CIO office circulated a similar memo. The working-group pattern is the procurement equivalent of "we read it, we will think about it, but we are not moving today." Whether any of these working groups produces a deployment-permitting recommendation in Q3 2026 will determine whether the disclosure has any procurement consequence inside US enterprise buyers.

The third pattern, present in approximately 3 of the 18 memos, treats the disclosure as procurement-actionable. The memos in this pattern instruct the technology team to evaluate V4 for non-production internal use cases — research, code generation in air-gapped environments, internal documentation summarisation that does not touch customer data — and to develop a deployment recommendation by a defined date. The memos that fall into this pattern come from buyers whose existing procurement posture was more permissive: two technology companies, one US logistics carrier. None of the three named Fortune 500 buyers in this pattern is in a regulated industry. The pattern is narrow. It is also the only pattern that opens a procurement path for V4 inside US enterprise deployment, and the procurement teams writing the memos are aware of that fact. The memos are precise about what they do and do not authorise. None of them authorises V4 in any production-customer-facing context.

The European procurement reaction is structurally different from the US reaction. Inside the EU, the buyer response has bifurcated along the EU AI Act compliance posture. Buyers operating under the EU AI Act general-purpose AI model obligations — which entered force on 2 August 2025 and which apply to systemic-risk models — have asked DeepSeek whether V4 will be classified as a systemic-risk model under the Act and whether DeepSeek will register the model with the European AI Office. DeepSeek's preliminary response, communicated in a 21 May 2026 statement, indicates that the company will register V4 with the AI Office and will comply with the Act's transparency and documentation requirements for systemic-risk models. That response opens a procurement path in the EU that is not yet open in the US. SAP's procurement office, BMW Group's IT office, and ING's risk function have all signalled internal openness to V4 evaluation conditional on the AI Office registration completing. The European procurement path is contingent. It is also more open than the US path on every dimension that procurement teams measure.

The UK procurement posture, mediated through the AI Safety Institute and the procurement guidance that Cabinet Office issued in March 2026, lands between the US and the EU positions. The AISI has signalled that it will conduct an independent evaluation of V4 in Q3 2026, and that the evaluation outcome will inform government procurement guidance. Until the AISI evaluation completes, the default UK government posture is non-procurement; the default UK private-sector posture is buyer-specific. HSBC's procurement office, the largest UK private-sector buyer reviewed in the 72-hour window, declined to commit to a position pending the AISI outcome. The position is sensible. It is also a hedge that defers the procurement decision into the second half of 2026.

The EU regulatory response from BSI and BMWi

The German Federal Office for Information Security — BSI — and the Federal Ministry for Economic Affairs and Climate Action — BMWi, now BMWK — issued a joint statement on 20 May 2026 acknowledging the V4 disclosure and committing to a technical review through the BSI's Center for AI Security. The statement is notable for what it does and does not say. It welcomes the disclosure as "a constructive contribution to the documentation standards expected under the EU AI Act systemic-risk provisions." It does not endorse the model, does not commit to a deployment authorisation, and does not modify the existing German federal IT acquisition guidance, which prohibits Chinese-origin software in critical infrastructure contexts under the IT Security Act amendments that came into force in 2024. The statement is procedurally significant — it is the first time a major EU member state regulator has formally responded to a Chinese-origin frontier model disclosure — and it is substantively conservative.

The European AI Office, the EU-level body established under the AI Act with responsibility for general-purpose AI model oversight, has not yet issued a formal response. The Office's director Lucilla Sioli has signalled in a 21 May 2026 press exchange that the Office will engage with DeepSeek on the registration process and will publish a guidance note within 60 days addressing the documentation standards exemplified by the V4 disclosure. The guidance note, when it ships, will be the most consequential European response. If the AI Office uses V4 as the benchmark for documentation completeness under the Act's systemic-risk provisions, every other frontier provider — OpenAI, Anthropic, Google DeepMind, Meta, Mistral, Cohere — will be assessed against a baseline that none of them currently meet at the same level of detail. That is the strategic consequence the DeepSeek disclosure has produced even before the Office responds. The frontier disclosure floor has been moved.

The French Direction Générale des Entreprises and the Autorité de Régulation des Communications Électroniques, des Postes et de la Distribution de la Presse have not issued comparable statements. France's posture on Chinese AI procurement is structurally less permissive than Germany's at the regulatory level but operates through procurement-channel restrictions rather than blanket prohibition. The French government's AI procurement guidance, updated on 1 April 2026, requires systemic-risk model deployments to undergo a CNIL data protection impact assessment, an ANSSI security review, and a strategic-autonomy review managed by the General Secretariat for Defence and National Security. The cumulative review surface for a Chinese-origin systemic-risk model under the French framework is substantial. The V4 disclosure does not remove any of the review surfaces; it provides better source material for the reviewers, which is a procedural improvement but not a procurement opening.

The Dutch Authority for Digital Infrastructure and the Belgian Centre for Cybersecurity have echoed the BSI position at a technical level and have indicated that they will participate in the European AI Office's V4 review through the formal coordination channels established under the AI Act. The Nordic regulators — Sweden's IMY, Denmark's Digitaliseringsstyrelsen, Norway's NSM — have signalled similar engagement. The Italian Agency for Digital Italy and the Spanish AESIA have remained quieter, with no formal statements as of 23 May 2026. The pattern across the EU regulator response is consistent: technical engagement with the disclosure, no procurement endorsement, deferral of any deployment-relevant guidance pending the European AI Office's coordinated review. The pattern is more open than the US procurement reaction and substantially more conservative than the disclosure itself would, on the surface, warrant. The conservatism is jurisdictional, not technical, and that is the same point the US procurement teams identified.

The Western reproducibility effort

The Hugging Face science team began a reproducibility evaluation of the V4 disclosure on 19 May 2026, hours after the publication. Hugging Face's head of science Leandro von Werra, working with the company's open-source evaluation cohort, published an initial verification report on 21 May 2026 covering the disclosed pre-training data composition, the tokeniser specification, and a subset of the benchmark results. The report verifies that the disclosed data composition is consistent with the model's observed behaviour on language-distribution probes, that the tokeniser specification reproduces a vocabulary closely matching the V4 release artifacts, and that 11 of 14 benchmark results fall within the standard variance band of independent re-runs. Three benchmark results — two on GPQA, one on a subset of SWE-Bench Verified — produced numbers slightly above DeepSeek's published figures, by a margin Hugging Face's team described as "within standard prompt-formatting variance" and "not indicative of a methodology issue."

ETH Zürich's Institute for Machine Learning, led by Andreas Krause, launched a parallel reproducibility effort on 20 May 2026 with a more architecturally focused scope. The ETH team is attempting to reproduce the V4 MoE routing protocol on a smaller-scale model — a 38-billion-parameter variant — using the disclosed routing specification and the disclosed auxiliary-loss-free balancing mechanism. The initial results, published in a preliminary blog post on 22 May 2026, indicate that the routing protocol behaves as DeepSeek describes on the smaller scale and that the disclosed expert-imbalance reduction reproduces within a 4 per cent margin of DeepSeek's reported figure. The ETH team has not yet attempted full-scale reproduction; the full-scale attempt would require compute that ETH does not currently have allocated for a single research project and would more plausibly proceed through a consortium with the EuroHPC compute facilities.

The Allen Institute for AI launched a third reproducibility effort focused on the post-training pipeline. AI2's Hanna Hajishirzi and Yejin Choi led an evaluation of the disclosed RLHF protocol against the AI2 Tulu training infrastructure, with the goal of assessing whether the V4 post-training methodology is consistent with the model's observed behavioural profile on instruction-following and refusal benchmarks. The preliminary AI2 report, expected by mid-June 2026, has not yet shipped, but Hajishirzi has indicated in conversation with at least three peer researchers that the constitutional library appendix in the V4 disclosure is "the most consequential single artifact" in the document because it provides the first public reference point for evaluating value-shaping consistency between an open-weights model's behaviour and its declared principles. The AI2 evaluation will be the most policy-relevant of the three reproducibility efforts when it ships.

The reproducibility effort taken together has produced a converging signal: the V4 disclosure is technically credible. The Hugging Face team's data and tokeniser verification, the ETH team's MoE routing verification, and the AI2 team's preliminary post-training engagement all indicate that the document represents an honest description of the training run. That convergence does not, by itself, modify the procurement posture in any of the buyer-side reactions described above. The procurement teams already assumed the disclosure was technically credible; the credibility was not the question. The question was whether technical credibility neutralises jurisdictional concerns about Chinese-origin frontier infrastructure, and the answer the procurement memos have produced is no. The disclosure raises the documentation floor for the entire frontier sector. It does not move the procurement boundary that US and most European buyers have drawn around Chinese-origin model deployment.

Does disclosure neutralise the China-origin concern?

The question DeepSeek's launch implicitly posed is whether unprecedented methodology disclosure neutralises the procurement concern about Chinese-origin AI infrastructure. The answer the first 72 hours have produced is a structured no. The procurement teams that reviewed the disclosure are not confused about its technical content. They acknowledge the credibility. They are declining to modify their procurement posture because the procurement posture is not, in their assessment, a question of methodological transparency. It is a question of jurisdictional risk, supply-chain risk, and the regulatory framework under which a Chinese-origin model would have to operate inside a Western enterprise. None of those dimensions changes when the training methodology is documented at higher resolution.

The jurisdictional risk dimension is the most cited in the procurement memos reviewed. The concern is that a Chinese-domiciled frontier provider operates under the Cyberspace Administration of China's data security framework, the Generative AI Services Management Provisional Measures, and the National Security Law. Under each of these instruments, the provider can be compelled to cooperate with Chinese state authorities on data access, model behavioural modification, or operational restriction. A US or European enterprise buyer deploying a Chinese-origin model in any environment that touches customer data faces a structural conflict between its own jurisdictional obligations — under HIPAA, GLBA, GDPR, the EU AI Act, sectoral financial regulation, or DOD cybersecurity requirements — and the obligations that the provider operates under in its home jurisdiction. The conflict does not disappear when the model is open-weights. It modulates only when the deployment surface is sufficiently isolated from customer data that the conflict has no practical contact surface, which is the niche that the more permissive procurement memos in the first 72 hours have identified.

The supply-chain risk dimension is the second most cited. Even an open-weights model deployment carries operational dependencies: the inference runtime, the tokeniser implementation, the safety-tuning checkpoints, the published documentation. Each of those dependencies originates from DeepSeek or from collaborators in DeepSeek's ecosystem. The supply-chain integrity of those dependencies is the assessment surface that the procurement teams treat as separable from the model weights themselves. The V4 disclosure does not directly address supply-chain integrity. It does not include the security review framework for DeepSeek's own development environment, the personnel security framework, the secure software development lifecycle controls, or the third-party audit attestations that Western enterprise buyers expect for any production-deployment vendor. Those omissions are conspicuous. They are the gap between methodology disclosure and procurement-grade vendor disclosure.

The regulatory framework dimension is the third. A US enterprise buyer deploying V4 must consider the application of the Bureau of Industry and Security's export controls on advanced AI, the Treasury's Office of Foreign Assets Control sanctions framework, the Federal Acquisition Regulation supplements that apply to federal contractors, and the sectoral regulations under which the buyer operates. None of these regulatory surfaces is addressed by the V4 disclosure. None of them softens in response to better methodology documentation. The Bureau of Industry and Security in particular has signalled in correspondence with at least four major US enterprise buyers that the disclosure does not modify the export control posture and that any US enterprise considering V4 deployment should treat the regulatory framework as set independently of the disclosure content. That signal landed in procurement offices within 48 hours and is doing more procurement work than the disclosure itself.

The deeper conclusion is that the open-weights conversation has been changed by the V4 disclosure, but it has not been changed in the direction DeepSeek's launch communications anticipated. The disclosure has raised the documentation floor for the frontier sector. It has not opened the procurement door. The two outcomes are not equivalent, and the procurement teams that read the disclosure are precise about the distinction. The conversation that has shifted is the conversation among Western frontier providers about what their own training documentation will need to look like by the end of 2026 to meet the floor that DeepSeek has just established. The conversation that has not shifted is the conversation inside Western enterprise procurement about whether Chinese-origin AI infrastructure can be deployed in production contexts. That conversation moves on jurisdictional and regulatory dimensions that DeepSeek's methodology disclosure does not touch.

What to watch

The V4 disclosure is five days old. The Q3 2026 cycle will determine whether it produces durable consequences for Western frontier disclosure norms and whether it opens any procurement paths inside European or US buyers.

• Whether the European AI Office's guidance note, expected within 60 days, treats the V4 disclosure as the documentation benchmark for systemic-risk model registration under the EU AI Act; if it does, OpenAI, Anthropic, Google DeepMind, Meta, Mistral, and Cohere will face a documentation-floor lift that none of them currently meets at the V4 level of detail, and the regulatory reporting requirements for the European market will rise materially in the Q3-Q4 window.
• Whether DeepSeek completes the European AI Office registration on the V4 model and what conditions the registration carries; the registration outcome will determine whether the procurement path that BMW, SAP, and ING have provisionally opened inside Europe converts into actual deployment authorisations, and whether the German BSI maintains or modifies its current conservative posture.
• Whether the UK AI Safety Institute's Q3 2026 evaluation of V4 produces a procurement-relevant outcome for the UK government and UK private-sector buyers; the AISI evaluation will be the single most consequential UK signal, and HSBC's procurement office and at least four other large UK private-sector buyers are explicitly waiting for it.
• Whether any US Fortune 500 buyer converts a working-group review into a production deployment authorisation; the 9 US procurement memos in the procurement-deferred pattern are the most consequential category — if any of them produces a Q3 2026 authorisation, the precedent will travel quickly across the US buyer cohort, and if none of them does, the procurement boundary will harden.
• Whether Anthropic, OpenAI, and Google DeepMind respond to the disclosure floor lift with their own expanded training documentation in the next two quarters; the strategic question is whether the V4 disclosure forces the Western frontier labs into a documentation arms race that none of them currently has the internal disclosure-readiness to win quickly, or whether the labs will hold their existing disclosure posture and accept the resulting regulatory pressure as the price of competitive secrecy.

Frequently asked

What is the most consequential single disclosure in the V4 document?

The data composition section and the constitutional library appendix are the two most consequential disclosures. The data composition section publishes per-source data origin at a granularity no previous frontier release has matched, with enumeration of 91 named domain-specific corpora. The constitutional library appendix publishes 188 named principles in English and Chinese, providing the first public reference point for evaluating the value-shaping architecture of a Chinese-origin frontier model. AI2's evaluation of the constitutional library against observed model behaviour is the analysis that policy teams across Europe and the US are most actively watching.

How does the compute disclosure compare to independent analyst estimates?

DeepSeek discloses 6.41 million H800 GPU-hours total for V4: 4.92 million for pre-training, 1.18 million for post-training, and 312,000 for ablations and evaluator-cohort experiments. SemiAnalysis and Epoch AI had estimated DeepSeek would require between 7.5 million and 8.2 million H800 GPU-hours for a V4-scale run, based on inferred parameter count and architectural scaling assumptions. The disclosed figure is roughly 19 per cent below the analyst estimate band. The figure sits inside the plausibility band when triangulated against the disclosed parameter count and MoE active-parameter ratio, and the reproducibility teams that have engaged with the disclosure have not flagged the compute number as inconsistent with the methodology.

Why has the US procurement reaction been more conservative than the European reaction?

The US procurement reaction is more conservative because the US regulatory framework around Chinese-origin AI infrastructure operates through a combination of export controls, sectoral regulations, and federal procurement guidance that does not modulate in response to methodology disclosure. The European reaction is more open because the EU AI Act provides a structured pathway — registration with the European AI Office, compliance with systemic-risk model obligations — that the V4 disclosure has positioned to use. The structural difference is regulatory architecture, not methodological assessment. The US procurement teams reading the disclosure acknowledge its technical credibility but do not have a regulatory framework that would translate technical credibility into a procurement path; the European procurement teams have such a framework, conditional on AI Office registration.

What is the European AI Office's role in the post-disclosure response?

The European AI Office is the EU-level body established under the AI Act with responsibility for general-purpose AI model oversight, including registration of systemic-risk models. The Office, led by Lucilla Sioli, will publish a guidance note within 60 days addressing the documentation standards exemplified by the V4 disclosure. If the guidance note uses V4 as the benchmark for documentation completeness under the systemic-risk provisions, every frontier provider operating in the European market — Anthropic, OpenAI, Google DeepMind, Meta, Mistral, Cohere — will need to lift its own documentation to match. The Office's posture toward DeepSeek's registration request will also determine whether procurement paths open inside European buyers; SAP, BMW, ING, and a number of public-sector buyers are explicitly contingent on the registration outcome.

Does the disclosure address supply-chain integrity concerns?

No. The V4 disclosure addresses training methodology, data composition, post-training protocols, and evaluation results. It does not address DeepSeek's secure software development lifecycle, personnel security framework, third-party audit attestations, security review of the development environment, or any of the operational integrity surfaces that Western enterprise buyers expect for production-deployment vendors. These omissions are the procurement gap between methodology disclosure and vendor-grade disclosure. Closing the gap would require a separate disclosure stream covering vendor security posture, which DeepSeek has not committed to publishing and which the Western buyer reaction has identified as the residual procurement concern even for buyers who treat the methodology disclosure as fully credible.

What is the most likely durable consequence of the V4 disclosure?

The most likely durable consequence is a documentation-floor lift across the entire frontier sector. DeepSeek has demonstrated that a frontier-scale training run can be documented at a level of detail that no Western lab has matched. Once the European AI Office references the V4 disclosure as the benchmark for systemic-risk model documentation — which the 60-day guidance note is widely expected to do — the regulatory pressure on Western labs to publish equivalent detail will rise materially. Anthropic, OpenAI, and Google DeepMind will have to choose between lifting their own training documentation to the V4 level, accepting the resulting regulatory friction in the European market, or attempting to lobby for a different documentation standard. The disclosure-floor change is the durable consequence; the procurement-boundary change is unlikely to be durable in any direction the disclosure intended.