Blog/EU AI Act and open-weight models
Regulation · November 2026

The EU AI Act and open-weight models: what a company self-hosting Qwen, DeepSeek or gpt-oss must do in 2026

By Samuel Seidel · Published November 7, 2026 · Sources checked November 7, 2026

If your company downloads an open-weight model such as Qwen, DeepSeek or gpt-oss and runs it for its own work, the EU AI Act's general-purpose model duties (Articles 53 to 55 of Regulation (EU) 2024/1689) sit with the model's developer, not with you. Your duties come from the AI system you build on it: the Article 50 transparency rules, which apply since 2 August 2026, the Article 5 prohibitions, and the high-risk rules only if the use is listed in Annex III, now from 2 December 2027 after the AI Omnibus, Regulation (EU) 2026/1744. You become a model provider yourself only if you place a modified model on the EU market and the modification used more than a third of the original training compute. This is my reading of the texts as of 7 November 2026, not legal advice.

DateWhat appliesLegal basis
2 Feb 2025Prohibited practices (Article 5) and AI literacy (Article 4)Article 113(a)
2 Aug 2025Obligations for providers of general-purpose AI (GPAI) models, Chapter VArticle 113(b)
2 Aug 2026Most of the Act, including Article 50 transparency; Commission can fine GPAI providersArticles 101, 113
2 Dec 2026New Article 5(1)(ba) and (bb) prohibitions; generative systems placed before 2 Aug 2026 must mark outputs2026/1744, new Articles 111(4) and 113
2 Aug 2027GPAI models placed on the market before 2 Aug 2025 must complyArticle 111(3)
2 Dec 2027High-risk obligations for Annex III uses (employment, credit, education, ...)2026/1744, Article 113 as amended
2 Aug 2028High-risk obligations for AI in regulated products (Annex I)2026/1744, Article 113 as amended

From the AI Act text on EUR-Lex, the Omnibus text (adopted 8 July 2026, published 24 July, in force 27 July) and the Commission's timeline page, updated 19 October 2026. All read 7 November 2026.

Are you a provider or a deployer?

For an off-the-shelf open model run in-house, you are a deployer of an AI system, and often also the provider of that system, but not the provider of the model. Article 3(3) defines a provider as whoever develops an AI system or GPAI model, or has it developed, and places it on the market or puts it into service under its own name. Article 3(4) makes a deployer anyone "using an AI system under its authority", except for personal non-professional use.

The Commission's guidelines on GPAI providers (published 18 July 2025, formally adopted as C(2025) 7719 on 19 November 2025) settle who the model provider is. Their examples in paragraph 51 of "placing on the market" include uploading a model "for the first time to a public catalogue, hub, or repository", offering it through an API, and copying it onto a customer's own infrastructure. When Alibaba uploads Qwen to Hugging Face, Alibaba is the provider. You download a model that is already on the market.

Two edge cases matter for self-hosters. First, a model counts as placed on the market when it is "used for internal processes that are essential for providing a product or service to third parties or that affect the rights of natural persons in the Union" (same paragraph). That clause is aimed at the model's own developer using it in-house, but it tells you how the Commission reads "internal use". Second, the guidelines set an indicative test for what a GPAI model is at all: training compute above 10^23 FLOP and the ability to generate text, audio, images or video (paragraph 17). By the 6 x parameters x tokens estimate used below, every current Qwen, DeepSeek or gpt-oss release clears that easily.

What do open-weight model providers owe, and what is exempt?

Article 53(1) lists four duties for every GPAI provider: (a) technical documentation for the AI Office, (b) information for downstream integrators, (c) a policy to comply with EU copyright law, and (d) a public summary of training content on the AI Office template. Article 53(2) removes (a) and (b) for models "released under a free and open-source licence" whose weights, architecture and usage information are public, and Article 54(6) removes the duty for non-EU providers to appoint an EU representative. Recital 104 is explicit that the copyright policy and training summary still apply.

The exemption has three limits:

Model familyWeights licenceFits the open-source exemption? (my reading)Code of Practice signatory?Source
Qwen3.6 / Qwen3.x (Alibaba)Apache 2.0Licence yes; compute not publishedNoHugging Face
DeepSeek-V4 (DeepSeek)MITLicence yes; compute not publishedNoHugging Face
gpt-oss (OpenAI)Apache 2.0Licence yesYes (OpenAI)Hugging Face
Llama 4 (Meta)Llama 4 Community LicenseNo: 700M MAU thresholdNoMeta licence

Signatory list from the Commission's Code of Practice page (updated 7 October 2026): 22 signatories including Amazon, Anthropic, Google, IBM, Microsoft, Mistral AI and OpenAI; xAI signed only the safety chapter. Alibaba, DeepSeek, Moonshot, Zhipu and Meta are not listed. Licence terms change by release; the planned open-weight licence comparison covers the other families.

The Code of Practice (published 10 July 2025) is voluntary. Its transparency and copyright chapters are a way to show compliance with Article 53, and the safety chapter applies only to systemic-risk models. A provider that has not signed must show compliance by other means. DeepSeek, for one, publishes a V4 "Technical Documentation" card (dated 27 April 2026) with sections on model properties, distribution, licences, intended uses and training data, and states the weights are MIT-licensed "as a free and open-source license". It gives no training compute or token count.

Does fine-tuning make you the provider? A worked calculation

Almost never, by the guidelines' own yardstick. Paragraph 60 says a downstream modifier becomes the provider of a GPAI model when "the training compute used for the modification is greater than a third of the training compute of the original model". If you cannot know the original figure, paragraph 61 substitutes a third of 10^23 FLOP (or a third of 10^25 for systemic-risk models). Even then, your obligations cover only the modification and its data (paragraph 65, recital 109). A typical company fine-tune of the kind described in the private fine-tuning guide is nowhere near the line.

Using the usual estimate of 6 x active parameters x training tokens:

Fine-tune Qwen3.6-35B-A3B (about 3B active) on 20,000 examples x 2,000 tokens x 3 epochs
  tokens  = 20,000 x 2,000 x 3      = 1.2e8
  compute = 6 x 3e9 x 1.2e8         = 2.2e18 FLOP

Fallback threshold (original compute unknown) = 1e23 / 3 = 3.3e22 FLOP
  3.3e22 / 2.2e18 = about 15,000 times more than this fine-tune

Original-compute route, Qwen3-235B-A22B (22B active, about 36T pretraining tokens)
  6 x 22e9 x 36e12 = 4.75e24 FLOP, one third = 1.6e24 FLOP
  a 1-billion-token continued pretraining run: 6 x 22e9 x 1e9 = 1.3e20 FLOP

The token count is from Qwen's Qwen3 announcement ("approximately 36 trillion tokens"); the rest is my arithmetic, pretraining only. The same formula shows how close current releases sit to the systemic-risk line: at 22B active parameters a model crosses 10^25 FLOP at about 76 trillion tokens, and at DeepSeek-V4-Pro's 49B active at about 34 trillion. DeepSeek does not publish V4's token count, so whether V4-Pro is presumed systemic-risk is for DeepSeek and the AI Office to establish, not for you to guess.

The trigger also needs a second step: placing the modified model on the market. A fine-tune kept on your own server for your own staff is not uploaded, sold or offered to anyone. If you publish it on a hub or sell API access to it in the EU and the modification crossed the one-third line, you are a provider of that model, with documentation duties limited to the modification.

What changed with the AI Omnibus in 2026?

Regulation (EU) 2026/1744 delayed the high-risk rules, softened AI literacy and left the GPAI obligations in Articles 53 to 55 unchanged; within Chapter V it only amended how codes of practice are monitored (Article 56(6)). The Commission proposed it on 19 November 2025, the co-legislators agreed on 7 May 2026, and it entered into force on 27 July 2026. The changes a self-hoster should know:

The GPAI timetable did not move. Since 2 August 2026 the Commission can enforce the model duties with fines of up to 3 percent of worldwide turnover or 15 million euros, whichever is higher (Article 101). Models already on the market before 2 August 2025 have until 2 August 2027 (Article 111(3)); the Commission's GPAI page repeats both dates.

What does a self-hosting company actually have to do?

Map each use case to a role, then apply Article 50 and check Annex III. For a typical internal deployment of Qwen, DeepSeek or gpt-oss, that comes to this list:

  1. Record your role per use case. Model provider: no, unless you publish or sell a model you modified with more than a third of its original training compute. System provider: yes, if you build the chatbot or tool and put it into service under your name. Deployer: yes, whenever staff use it at work. Note that the open-source exclusion for AI systems in Article 2(12) does not cover systems that fall under Article 50.
  2. Tell people they are talking to an AI. Article 50(1), unless it is obvious to a reasonably well-informed person. Applies since 2 August 2026.
  3. Mark generated content. Article 50(2) requires the provider of a system that generates synthetic text, images, audio or video to have outputs "marked in a machine-readable format and detectable as artificially generated". That can include an internal tool you built and put into service, so read Article 50(2) in full, including its exceptions, before assuming it does not apply.
  4. Disclose deepfakes and AI-written public-interest text. Article 50(4), as a deployer, unless a person with editorial responsibility reviewed the text.
  5. Check Annex III before December 2027. CV screening, credit scoring, exam grading and similar uses make the system high-risk. Deployers then owe Article 26 duties: follow the instructions, assign competent human oversight, keep automatically generated logs for at least six months, and inform workers before workplace use. Changing a system's purpose so it becomes high-risk can make you its provider under Article 25(1).
  6. Keep the upstream paperwork. Because open-source providers are exempt from Article 53(1)(b), you may get less integration information than from a closed vendor. File the model card, licence and training-content summary with the date you read them, and record which checkpoint revision you run; the model versioning guide covers that.
  7. Train the people who use it. Article 4 as amended is a best-efforts duty, but you should be able to show what you did.

GDPR is a separate question with its own answers; GPUwerk's compliance page explains what changes when the model runs on dedicated hardware and what does not (the Article 28 DPA is published), and the audit logging guide covers keeping the kind of logs Article 26(6) asks for. The hosting choice does not change your AI Act role; the use case does.

Is it safer to use a closed API instead?

Not under the AI Act. Your Article 50 and Article 26 duties are the same whether the model sits behind OpenAI's API or on your own machine, because they attach to the system and its use. What differs is the supplier: a closed-model provider owes you Article 53(1)(b) integration information and has no open-source exemption, while an open-weight provider may skip it. My view: for an internal assistant on a 2026 open model with an Apache 2.0 or MIT licence, the AI Act is a manageable checklist rather than a reason to avoid self-hosting. The harder questions are data protection and whether the use lands in Annex III. For those, a lawyer who knows your use case is worth more than any article, including this one. If you want to compare hosting routes, the private LLM hosting page sets out the infrastructure side.

FAQ

Does the EU AI Act apply to open-source LLMs?

Yes, with exemptions. Under Article 53(2) of Regulation (EU) 2024/1689, providers of general-purpose models released under a free and open-source licence with public weights skip the technical documentation duties in Article 53(1)(a) and (b) and, under Article 54(6), the EU representative. They still need a copyright policy and a public training-data summary, and the exemption ends if the model has systemic risk, presumed above 10^25 FLOP of training compute.

Am I a provider if I self-host an open-weight model?

Usually not a provider of the model. The Commission's GPAI guidelines say the developer that first makes the model available on the EU market, for example by uploading it to a hub, is the provider. You become a model provider if you modify it with more than a third of its original training compute and place it on the market. You can still be the provider of the AI system you build on it.

What changed in the EU AI Act in 2026?

Regulation (EU) 2026/1744, the AI Omnibus, entered into force on 27 July 2026. It moved Annex III high-risk obligations from 2 August 2026 to 2 December 2027 and Annex I product rules to 2 August 2028, softened the Article 4 AI literacy duty, added two prohibitions from 2 December 2026 and gave older generative systems until 2 December 2026 to mark outputs.

What applies from 2 August 2026 under the AI Act?

Most of the Regulation, including the Article 50 transparency duties: telling people they are talking to an AI system, machine-readable marking of generated content, and disclosing deepfakes. From the same date the Commission can fine providers of general-purpose AI models up to 3 percent of worldwide turnover or 15 million euros under Article 101. Annex III high-risk rules were deferred to December 2027.

Do Qwen and DeepSeek comply with the EU AI Act?

That is their providers' obligation, not yours, and the record is thin. Neither Alibaba nor DeepSeek appears among the 22 signatories of the GPAI Code of Practice listed by the Commission on 7 November 2026. Both release weights under licences (Apache 2.0, MIT) that fit the open-source exemption. DeepSeek publishes a V4 technical documentation card dated 27 April 2026.

Sources

Every figure above is attributed to its source in the sentence or table where it appears. Pages were read on 7 November 2026 unless a date says otherwise.

Related pages

First top-up: pay $20, get $30 in credit

Keep the model on hardware only you use.

A dedicated DGX Spark in the Czech Republic does not change your AI Act role, but it gives a short answer to who can reach the data while the model runs.

A dedicated DGX Spark in the EU, $0.79/hour, billed per minute. $30 of credit is about 37 hours of compute.

Create account What GPUwerk documents for GDPR and the AI Act