---
title: "Open Weights Don’t Automatically Mean Open Source for AI Models – IOWIS"
description: "Open weights do not automatically make an AI model open source. The article explains the role of model weights, training data, code and licences."
source: "https://iowis.com/en/blog/open-weights-vs-open-source"
language: "en"
---

Source: [Open Weights Don’t Automatically Mean Open Source for AI Models – IOWIS](https://iowis.com/en/blog/open-weights-vs-open-source)

[Back to articles](https://iowis.com/en/blog)

AI in Business

# Open Weights Don’t Automatically Mean Open Source

Why published model weights do not yet tell us how open an AI model really is

October 01, 2026·8 Minuten Reading time

![Black kettlebells on a gym floor, with a large kettlebell in the foreground.](https://iowis.com/media/Openweights_bild.jpg)

Photo: [Jonathan Borba / Pexels](https://www.pexels.com/photo/a-kettlebell-and-a-barbell-on-a-gym-floor-27810155/), edited.

With open-weight models, the trained weights are accessible. Code, training data and the development process can still be disclosed to very different degrees.

## What “Open” Actually Means in AI

A model provider can make hundreds of gigabytes of an AI model’s weights publicly available for download and still reveal very little about how the model was created. Others also publish training code, data sources, configurations and intermediate training checkpoints for their models. In discussions, both types of release are quickly grouped under the umbrella term “open models”.

This can now be seen clearly with well-known model families such as Kimi, Qwen, gpt-oss, Llama or Olmo. Model developers use terms such as “open weight”, “open model”, “open source” or “fully open”, but they do not always mean the same thing by them.

There is a technical reason for this inconsistency. With AI, it is genuinely harder to determine what corresponds to the “source” in open source. A trained model consists of several components, each of which reveals something different about how it works and how it came to be.

## What Model Weights Actually Are

During the training of a language model, a very large number of numerical parameters are adjusted so that the model learns statistical relationships in the training data. The result of this process is the model weights.

The weights represent the trained state of the model. Together with the necessary architecture and inference software, the model can be run on suitable hardware. The weights can be analysed, trained further or used as a starting point for fine-tuning, provided the applicable terms of use allow it.

How this result came about depends on other components: the model architecture, the training code, the selection and preparation of the data, the training parameters, the evaluation methods used, and a large number of other decisions made during development.

This is precisely where a trained neural network differs from conventional software. In a program, the source code directly describes the logic implemented by the developers. In a language model, a substantial part of its eventual behaviour was not directly programmed but learned during training.

The weights encode this learned behaviour. They are not a human-readable account of how it came about. You cannot simply read from them which specific data a model was trained on, how those data were weighted, or which training decisions led to a particular capability or weakness.

That is why publishing the weights alone is not sufficient under a broader notion of open source for AI.

## What Can Be Made Available Beyond the Weights

A release of a trained model can include other parts of the development process in addition to the weights. These can include training and inference code, information about the training data — for example, their provenance, selection, filtering and mixing — as well as configurations, hyperparameters, evaluation methods and intermediate stages of training.

The current [Open Source AI Definition 1.0 from the Open Source Initiative](https://opensource.org/ai/open-source-ai-definition) sets out this idea in relatively concrete terms. An open-source AI system should be available for anyone to use, study, modify and share for any purpose. To enable this, the definition requires the relevant code in addition to the model parameters, as well as sufficiently detailed information about the training data and how they were processed. The original raw data do not necessarily have to be provided in full, but the documentation should go far enough to allow a skilled third party to develop a substantially equivalent system.

[In its report](https://www.oecd.org/en/publications/benefits-of-ai-openness_746e8c9a-en.html) [Benefits of AI openness](https://www.oecd.org/en/publications/benefits-of-ai-openness_746e8c9a-en.html)[, the OECD](https://www.oecd.org/en/publications/benefits-of-ai-openness_746e8c9a-en.html) considers an even broader set of components in the AI ecosystem. It distinguishes, among other things, compute infrastructure, data, models, software and tooling, and applications. Depending on the layer, openness can therefore refer to different things: for models, for example, accessible weights and documentation; for data, their availability and provenance; for software, open code; and for technical infrastructure, open standards and interfaces. The report therefore frames openness in terms of which parts of an AI system are accessible and reusable.

The two approaches place emphasis on different aspects. What they have in common is that “open” in AI can describe more than the availability of trained weights.

## How Open-Weight Model Releases Differ

The term “open weight” refers to something fairly specific: the trained weights of a model are made publicly available. How much is released beyond the weights varies from model to model.

For [gpt-oss](https://help.openai.com/en/articles/11870455-openai-open-weight-models), OpenAI provides reference implementations for inference in addition to the weights. The models are released under [Apache 2.0](https://github.com/openai/gpt-oss/blob/main/LICENSE); the [gpt-oss usage policy](https://github.com/openai/gpt-oss/blob/main/USAGE_POLICY) also applies. The term “open-weight” is deliberately chosen here. It describes the publicly available model state without describing the entire training process as open.

[Kimi K3](https://huggingface.co/moonshotai/Kimi-K3) is likewise released by Moonshot as an open-weight model. In addition to the weights, code and detailed technical information about the model are available. Its use is governed by a dedicated Kimi licence, which generally permits broad use and modification but imposes additional conditions on certain commercial offerings.

Qwen3.8 shows that different conditions can apply even within the same model family. [Qwen3.8-27B](https://huggingface.co/Qwen/Qwen3.8-27B) is released under Apache 2.0. [Qwen3.8-Flash-Next](https://huggingface.co/Qwen/Qwen3.8-Flash-Next), by contrast, uses the Qwen Community License 1.0. It likewise permits broad use but includes additional requirements for certain commercial services.

Taken together, these examples show the range of what model developers release under the open-weight label. In each case, the weights are accessible; the scope of the accompanying materials and the terms of use differ.

## “Open Source” Is Not Used Consistently Either

Additional ambiguity arises when model developers themselves use the term open source.

Meta describes [Llama 4](https://ai.meta.com/resources/models-and-libraries/llama-downloads/) as open source in its communications. The weights are accessible, and the Llama Community License allows broad use, modification and redistribution. However, it also contains restrictions that do not align with the OSI’s Open Source Definition. These include usage restrictions in the incorporated Acceptable Use Policy and a requirement for certain very large companies to obtain a separate licence from Meta.

For the OSI, restrictions of exactly this kind matter. Its Open Source Definition requires, among other things, that a system may be used for any purpose and that the licence does not exclude particular fields of use.

That is precisely why the question “Is Llama open source?” only gets us so far. It conflates a label used by the model developer, a specific licence and differing views about which components of an AI model must be published for it to be genuinely open.

## When the Training Process Is Open Too

The [Olmo project from the Allen Institute for AI (Ai2)](https://allenai.org/olmo), a nonprofit AI research institute, shows how far a release can go beyond open weights.

For Olmo, the release includes final model weights, training code, data or extensive data artefacts, configurations, evaluations, and numerous checkpoints from different stages of training.

Those checkpoints in particular change what can be investigated about a model.

With a release of the final weights, it is possible to analyse how the trained model responds to particular inputs and which capabilities or weaknesses it exhibits. Intermediate training checkpoints also make it possible to observe how such properties develop. Researchers can, for example, compare different points in training or investigate when particular information was learned and whether it was later lost again.

Ai2 has used the open Olmo checkpoints, among other things, to study how facts are acquired during pretraining and how stable this knowledge remains over the subsequent course of training.

For research, what matters most is that claims about how the model came to be can be checked more readily. A full training run from scratch remains expensive and technically demanding even when data, code and checkpoints are openly available.

Olmo is not an isolated example. [Apertus](https://www.apertus-ai.org/) from the Swiss AI Initiative also provides weights, code, extensive information, and tools relating to the training data. Earlier research projects such as Pythia from EleutherAI or LLM360 have likewise published data, code and intermediate checkpoints.

Projects like these also make the difference from a weights-only release more tangible: alongside the final state of the model, parts of its development process can also be studied.

## The EU AI Act Draws Its Own Distinction

The [EU AI Act](https://eur-lex.europa.eu/legal-content/EN/TXT/?uri=CELEX%3A02024R1689-20260727) also uses the concept of open or open-source AI models for a specific regulatory purpose.

Providers of certain general-purpose AI models must prepare extensive technical documentation and provide information to providers that integrate their models into their own AI systems. For certain models, [Article 53](https://ai-act-service-desk.ec.europa.eu/en/ai-act/article-53) provides exemptions from some of these obligations when they are released under a free and open-source licence.

For this exemption, the model parameters, including the weights, as well as information about the architecture and model usage must also be publicly available.

Other obligations remain. These include, in particular, a sufficiently detailed public summary of the content used for training. The open-source exemptions described above also do not apply to general-purpose AI models with systemic risk.

The EU AI Act therefore sets conditions for a specific regulatory exemption. Its criteria need not coincide with the OSI’s Open Source AI Definition or the terminology used by individual model developers.

## Why Open Source Is Harder to Define for AI

The debate around open weight and open source ultimately stems from a particular characteristic of machine learning.

The term “open weight” precisely describes the availability of one specific component: the trained model weights. These can already enable a wide range of practical uses even when comparatively little has been published about the original training process.

Open source asks a broader question in AI. It concerns which information and components must be accessible for a system to be meaningfully studied, modified and further developed independently. Projects such as Olmo show how far such openness can go; other releases deliberately focus more narrowly on the weights.

With conventional software, the source code brings together much of what is needed to study and modify the software. With trained models, this information is distributed across several parts of the development process.

That is why open weight is not simply a weaker form of open source. It is a precise statement about one particular part of an AI model. The broader open-source debate begins with the question of how much of the development process must be accessible for the trained model itself to be open to further development.

That is what makes this debate distinctive: the concept of “source”, inherited from the software world, maps only imperfectly onto trained models.

[Sovereign AI for businesses](https://iowis.com/en/sovereign-ai)

[Back to articles](https://iowis.com/en/blog)

## Structured data

```json
[
  {
    "@context": "https://schema.org",
    "@graph": [
      {
        "@id": "https://iowis.com/#organization",
        "@type": "Organization",
        "name": "IOWIS",
        "url": "https://iowis.com/",
        "description": "Local, on-premise AI for organisations that handle sensitive data. Keep full data sovereignty, meet GDPR requirements and avoid cloud dependencies.",
        "sameAs": [
          "https://www.linkedin.com/company/iowis",
          "https://www.instagram.com/iowis_com/"
        ],
        "address": {
          "@type": "PostalAddress",
          "streetAddress": "IOWIS GmbH",
          "postalCode": "74889",
          "addressLocality": "Sinsheim",
          "addressCountry": "Germany"
        },
        "foundingDate": "2012"
      },
      {
        "@id": "https://iowis.com/#website",
        "@type": "WebSite",
        "name": "IOWIS",
        "url": "https://iowis.com/",
        "description": "Local, on-premise AI for organisations that handle sensitive data. Keep full data sovereignty, meet GDPR requirements and avoid cloud dependencies.",
        "publisher": {
          "@id": "https://iowis.com/#organization"
        }
      },
      {
        "@id": "https://iowis.com/en/blog/open-weights-vs-open-source#webpage",
        "@type": "WebPage",
        "url": "https://iowis.com/en/blog/open-weights-vs-open-source",
        "name": "Open Weights Don’t Automatically Mean Open Source for AI Models",
        "description": "Open weights do not automatically make an AI model open source. The article explains the role of model weights, training data, code and licences.",
        "isPartOf": {
          "@id": "https://iowis.com/#website"
        },
        "about": {
          "@id": "https://iowis.com/#organization"
        },
        "publisher": {
          "@id": "https://iowis.com/#organization"
        }
      },
      {
        "@type": "BlogPosting",
        "headline": "Open Weights Don’t Automatically Mean Open Source",
        "url": "https://iowis.com/en/blog/open-weights-vs-open-source",
        "datePublished": "2026-10-01T12:00:00.000Z",
        "dateModified": "2026-10-01T13:40:43.216Z",
        "keywords": [
          "AI in Business"
        ],
        "publisher": {
          "@id": "https://iowis.com/#organization"
        },
        "image": {
          "@type": "ImageObject",
          "url": "https://iowis.com/api/media/file/Openweights_bild.jpg",
          "height": 877,
          "width": 1314
        }
      }
    ]
  }
]
```