Skip to content
ContentLora

    Tip: press / anywhere to search.

    Explainer

    Open-weight vs closed AI models, explained

    An open-weight AI model is one whose learned parameters, or weights, anyone can download and run, while many developers give limited or no public access to their models' weights.[1][2] "Open weights" is not the same as open source: the Open Source Initiative's definition also requires freedom to use the model for any purpose and enough information about the training data to rebuild it.[3][4]

    Editor reviewedUpdated Explainer20 claims · 100% primaryArtificial intelligence, Frontier AI, Tech policy
    Reading level: 3 passages differ; the rest is the same for everyone.

    This explainer covers what “open” means for AI models, the main licence types, well-known examples and the policy debate. It does not rank models or recommend one. Choose a reading level in each section.

    What weights are

    A trained AI model is a large set of numbers, called weights, that the model learned during training. Those weights, plus the model’s design and the code that runs it, are the model.[1] If a developer lets anyone download the weights, the model is “open-weight”. If the weights stay on the developer’s servers and you can only send it questions through an app or API, it is “closed”.[2]

    The Open Source Initiative (OSI) defines an AI model as its architecture, parameters (including weights) and inference code, and weights as the learned parameters that overlay the architecture to map inputs to outputs.[1] NTIA’s 2024 report on “dual-use foundation models with widely available model weights” contrasted developers that give limited or no public access to their weights with those, including Meta, Google, Microsoft and Mistral, that have released models with widely available weights.[2]

    Open weights vs open source

    “Open source” has a stricter meaning. The OSI says an open-source AI system must let anyone use it for any purpose without asking, study how it works, change it and share it.[3] It also asks for enough detail about the training data that a skilled person could build a similar system.[4] Many models with downloadable weights do not meet that bar.

    The OSI’s Open Source AI Definition 1.0 grants four freedoms (use, study, modify, share) and applies them both to whole systems and to components such as weights.[3] Its “preferred form to make modifications” requires data information sufficient for a skilled person to build a substantially equivalent system, alongside the full training and inference code.[4] In February 2025 the OSI said Meta’s Llama licences still fail the Open Source Definition, citing restrictions on users and fields of endeavour.[5]

    Licences and examples

    Meta says it has released its Llama models openly while also training closed models.[6] The Llama 4 Community License requires any licensee with more than 700 million monthly active users to request a separate licence, which Meta may refuse.[7]

    DeepSeek-R1‘s code and weights are under the MIT License, which DeepSeek says allows commercial use and distillation into other models.[8] Alibaba’s Qwen team licenses all its open-weight models under Apache 2.0.[9] Mistral AI mixes licences: it lists Mistral Small 4 under Apache 2.0 and Mistral Large 4 under its own Mistral Ecosystem Open Weight License.[10]

    xAI released Grok-1, a 314-billion-parameter mixture-of-experts model, under Apache 2.0 in March 2024.[11] On 5 August 2025 OpenAI released gpt-oss-120b and gpt-oss-20b under Apache 2.0.[12] OpenAI says the larger model has 117 billion total parameters and runs on a single 80 GB GPU.[13]

    The safety debate

    NTIA’s 2024 report said widely available weights could create risks to national security, safety, privacy and civil rights.[14] Before releasing gpt-oss, OpenAI also tested an adversarially fine-tuned version of the larger model.[15]

    NTIA listed affirmative misuse, failures of oversight and unclear accountability among the risks of releasing certain foundation-model weights.[14] OpenAI evaluated an adversarially fine-tuned gpt-oss-120b under its Preparedness Framework before release.[15] The US Center for AI Standards and Innovation (CAISI) reported in September 2025 that DeepSeek’s R1-0528 model was 12 times more likely than the US frontier models it tested to follow malicious instructions.[16] In September 2026 it called Z.ai’s GLM-5.3 the most cyber-capable open-weight model released to date.[17]

    Competition and policy

    NTIA said open foundation models widen who can take part in AI development, decentralise market control from a few large developers and let users avoid sending data to third parties.[14] It recommended that the US government monitor risks and be ready to act, rather than restrict open weights at that time.[18] The White House’s July 2025 AI Action Plan said open models help organisations with sensitive data and academic researchers, have geostrategic value, and that America needs leading open models.[19]

    In Europe, the AI Act exempts general-purpose models released under a free and open-source licence with public weights from some documentation duties, but not models with systemic risk.[20]

    Next in the crash course

    How AI image and video generators work

    How diffusion models and video transformers make images and clips, where training data comes from, and how watermarks, C2PA and EU labelling rules work.

    More in AI

    Questions readers ask

    What are model weights?

    Weights are the learned parameters that sit on top of a model's architecture and turn an input into an output. Together with the architecture and inference code, they make up the model.[1]

    Is Llama open source?

    Meta describes its Llama releases as open, but the Open Source Initiative says the Llama licences fail the Open Source Definition because they restrict users and uses. The Llama 4 licence, for example, requires companies with more than 700 million monthly active users to ask Meta for a separate licence.[6][5][7]

    Has OpenAI released an open-weight model?

    Yes. On 5 August 2025 OpenAI released gpt-oss-120b and gpt-oss-20b under the Apache 2.0 licence.[12]

    Does the EU AI Act treat open-source models differently?

    Partly. Providers of general-purpose models released under a free and open-source licence with public weights are exempt from some documentation duties, but the exemption does not apply to models with systemic risk.[20]

    Did the US government decide to restrict open-weight models?

    No. In July 2024 NTIA recommended monitoring risks rather than restricting open weights, and the July 2025 AI Action Plan said America needs leading open models.[18][19]

    Sources

    · 20 claims from 12 publishers

    Each numbered claim is a statement we checked against the sources listed with it. Tap a number in the text to see its sources in place.

    1. [1]

      The OSI definition describes an AI model as its architecture, parameters (including weights) and inference code, and defines weights as the learned parameters that produce an output from a given input. confirmedas of 2026-11-07

    2. [2]

      NTIA's July 2024 report noted that many developers give limited or no public access to their advanced models' weights, while others, including Meta, Google, Microsoft and Mistral, have released models with widely available weights. confirmedas of 2024-07-30

    3. [3]

      The Open Source Initiative's Open Source AI Definition (version 1.0) says an Open Source AI system must be made available on terms that let anyone use it for any purpose without asking permission, study how it works, modify it and share it. confirmedas of 2026-11-07

    4. [4]

      Beyond weights, the OSI definition requires "data information" detailed enough for a skilled person to build a substantially equivalent system, plus the code used to train and run it. confirmedas of 2026-11-07

      • The Open Source AI Definition 1.0 · Open Source Initiative · Primary source · Preferred form to make modifications to machine-learning systems (retrieved 2026-11-07)
      • The Open Source AI Definition 1.0 · Open Source Initiative · Primary source · Preferred form to make modifications to machine-learning systems (retrieved 2026-11-07)
    5. [5]

      In February 2025 the OSI said Meta's Llama licences fail the Open Source Definition because they restrict who may use the models and for what purposes. confirmedas of 2025-02-18

    6. [6]

      Meta's 2025 annual report says it has released its Llama foundation models openly, has not released everything it develops, and expects to train a mix of open and closed models. confirmedas of 2026-01-29

    7. [7]

      The Llama 4 Community License requires any licensee whose products had more than 700 million monthly active users to request a separate licence from Meta, which Meta may grant at its sole discretion. confirmedas of 2025-04-05

    8. [8]

      DeepSeek licenses the DeepSeek-R1 code and model weights under the MIT License, allowing commercial use, modification and distillation into other models. confirmedas of 2026-11-07

    9. [9]

      Alibaba's Qwen team says all its open-weight Qwen models are licensed under Apache 2.0. confirmedas of 2026-11-07

      • Qwen3 repository README · Qwen team, Alibaba Cloud (GitHub) · Primary source · License Agreement (retrieved 2026-11-07)
    10. [10]

      As of November 2026 Mistral AI lists Mistral Small 4 under Apache 2.0 and Mistral Large 4 under its own Mistral Ecosystem Open Weight License. confirmedas of 2026-11-07

      • Models overview · Mistral AI · Primary source · Models table (name, licence and description cells; checked in the saved page 2026-11-14) (retrieved 2026-11-07)
      • Models overview · Mistral AI · Primary source · Models table (name, licence and description cells; checked in the saved page 2026-11-14) (retrieved 2026-11-07)
    11. [11]

      On 17 March 2024 xAI released the base-model weights and architecture of Grok-1, a 314-billion-parameter mixture-of-experts model, under the Apache 2.0 licence. confirmedas of 2024-03-17

    12. [12]

      On 5 August 2025 OpenAI released gpt-oss-120b and gpt-oss-20b, two open-weight language models, under the Apache 2.0 licence. confirmedas of 2025-08-05

      • Introducing gpt-oss · OpenAI · Primary source · 2025-08-05 · Introduction (hand-saved copy; fetch-source got HTTP 403) (retrieved 2026-11-07)
    13. [13]

      OpenAI says gpt-oss-120b has 117 billion total parameters and gpt-oss-20b has 21 billion, and that the larger model runs on a single 80 GB GPU. confirmedas of 2025-08-05

      • Introducing gpt-oss · OpenAI · Primary source · 2025-08-05 · Pre-training & model architecture (retrieved 2026-11-07)
      • Introducing gpt-oss · OpenAI · Primary source · 2025-08-05 · Introduction (retrieved 2026-11-07)
    14. [14]

      NTIA's report said open foundation models widen participation in AI development, decentralise market control and let users avoid sharing data with third parties, but that widely available weights could also create risks to national security, equity, safety, privacy and civil rights through misuse, failures of oversight or a lack of clear accountability. confirmedas of 2024-07-30

    15. [15]

      Before releasing gpt-oss, OpenAI tested an adversarially fine-tuned version of gpt-oss-120b under its Preparedness Framework. confirmedas of 2025-08-05

      • Introducing gpt-oss · OpenAI · Primary source · 2025-08-05 · Safety & worst-case fine-tuning (retrieved 2026-11-07)
    16. [16]

      In September 2025 CAISI published an evaluation of three DeepSeek models against four US models across 19 benchmarks, finding the DeepSeek R1-0528 model 12 times more likely to follow malicious instructions than the US frontier models evaluated. confirmedas of 2025-09-30

    17. [17]

      CAISI assessed Z.ai's GLM-5.3 in September 2026 as the most cyber-capable open-weight model released to date, while lagging the US frontier by about four months on its cyber benchmarks. confirmedas of 2026-09-17

    18. [18]

      Rather than restricting open weights, NTIA recommended in July 2024 that the US government monitor risks and be ready to act, while supporting developers' choice to release weights at that time. confirmedas of 2024-07-30

    19. [19]

      The White House's July 2025 AI Action Plan said open-source and open-weight models help startups and organisations with sensitive data, are essential for academic research, and have geostrategic value, and that America needs leading open models. confirmedas of 2025-07-23

    20. [20]

      The EU AI Act exempts providers of general-purpose AI models released under a free and open-source licence, with public weights, from some documentation duties, but not if the model has systemic risk. confirmedas of 2024-07-12

    Revision history (1)
    1. Page created.

    Created Nov 7, 2026. Last reviewed by an editor on Nov 8, 2026. Next scheduled review: Mar 7, 2027.

    Cite this page

    "Open-weight vs closed AI models, explained." ContentLora, updated Nov 7, 2026. https://contentlora.com/explain/open-weight-vs-closed-models-explained

    Spotted an error? Suggest a correction or emailcorrections@contentlora.com.