How open source brought AI closer to home

The story of shared software and what it means to run an AI on your own computer

BunCosmos | 26 October 2026

Imagine opening an AI assistant on your laptop, asking it to explain a difficult paragraph, and receiving an answer without sending that paragraph to an AI company’s servers. The model is running on your machine. You chose which one to download, and you can decide when to replace it.

That is one possibility offered by local AI. Behind it sits a much older story: people sharing software, giving others permission to change it, and building useful things from one another’s work.

This is the world usually called open source. It includes far more than AI, from operating systems to animation software. Understanding it means separating three questions that often get mixed together: What are you allowed to change? Where does the software run? And how capable is it?

First imagine receiving the recipe

When you use an ordinary application, you usually see its buttons, menus and results. Its source code is the human-readable set of instructions behind those things. Think of the application as a finished cake and the source code as its recipe.

Open-source software comes with permission to inspect that recipe, use it, change it and redistribute it under its licence. A licence is the legal document explaining those rights and their conditions. Merely putting code on a public website does not grant all of them. [1]

That matters because “free to download” and “open source” answer different questions. A free app can keep its code secret. Open-source software can be sold. In the related phrase “free software,” the central idea is freedom to use and change software, rather than a promise that every product or service costs nothing. [2]

You do not have to become a programmer to benefit. If a tool stops meeting your needs, the permission to change it can allow another developer or company to maintain it or create an alternative. That is a possibility, however, not a guarantee that someone will do the work for you.

The idea came before the name

Software sharing existed in early computing communities long before anyone called it open source. The movement did not begin with a single website or one generous company. It grew through overlapping traditions of research, collaboration and arguments about who should control the tools people use. [4]

A major turning point came on 27 September 1983, when programmer Richard Stallman announced GNU, a project to build a free operating system. Work began in January 1984. Its ambition was practical and political: users should be able to study, share and improve their software. [3]

In 1991, Finnish student Linus Torvalds began Linux. Technically, Linux is a kernel: the central part of an operating system that manages hardware and helps programs run. Combined with other software, including GNU tools, it became the basis of complete operating systems. [5]

The name “open source” arrived later. It was adopted at a meeting on 3 February 1998, with Christine Peterson credited with suggesting the term. Eric Raymond and Bruce Perens subsequently founded the Open Source Initiative. The new language helped explain collaborative software development to a wider business audience. It named an approach whose roots were already much older. [4]

How strangers build something together

A project commonly keeps its files and their change history in a repository, often on a service such as GitHub. Someone reports a problem. Another person proposes a fix. Maintainers, the people responsible for the project, review proposed changes and decide what enters the official version. [6]

This is more organised than letting the entire internet edit one file. People can suggest changes without gaining control of the project. They can also help by testing, translating instructions or making documentation clearer. A useful contribution does not have to be a thousand lines of code. [6]

If people want to take a project in a different direction, an appropriate licence allows them to make a separate version, called a fork. The original project can continue alongside it. Having that option is one reason licensing matters as much as the ability to look at the files. [1, 6]

Projects still need leadership, money and maintenance. Some depend heavily on volunteers; others have contributors paid by companies. Businesses can charge for hosting, support, security work or customised versions. Red Hat is a familiar example of a company building a commercial service around open-source software. The code’s availability does not remove the cost of keeping important systems dependable. [7]

What is available and how much is there

The range is broad. There are operating-system components, browsers, creative applications, tools for building websites, databases and software for running AI. [8] The AI world also shares model files, datasets and demonstration apps, although each of those can have different terms. [17, 18] It is an ecosystem, rather than one product called “Open Source.”

There is no dependable single number for all open-source projects worldwide. What would we count: an original project, each separate version, a copy made for experimentation, or every small component inside a larger application? Different directories measure different things.

GitHub’s 2025 Octoverse report counted 395 million public repositories. [8] That demonstrates the scale of public code sharing. It does not establish that there were 395 million distinct, properly licensed open-source projects: a public repository is a storage and collaboration unit, not a certificate of open-source status. [2]

For AI, Hugging Face’s public catalogue displayed more than three million model repositories when checked on 26 October 2026. That includes adapted models and versions prepared in different formats. [9] It does not mean three million independently invented ChatGPT equivalents, nor that every listed model qualifies as open source. [18]

AI adds another layer to the recipe

An AI model is not made only from instructions written by programmers. Training also produces large collections of numbers called weights. Those numbers encode patterns learned during training and help determine the model’s output. To run a downloaded model, suitable software loads its weights and performs the calculations needed to generate an answer. [12, 21]

This is why AI makes the word “open” more complicated. A developer might publish the software used to run a model without publishing its weights. Another might release the weights while keeping important details of training unavailable. A downloadable model is often described more precisely as “open weight.” [10, 14] Its licence still needs checking. [18]

The Open Source Initiative’s Open Source AI Definition 1.0 asks for more than weights: it requires relevant code, model parameters and sufficiently detailed information about the training data to support study and modification. It does not require every raw training example to be publicly downloadable. This is an influential definition, not a reason to assume every vendor using “open” meets it. [10]

Those distinctions let us describe a system honestly. An open-weight model might run on a rented cloud server. A local application might contain proprietary software. One system can combine open-source tools, a model with its own licence and a paid interface. The labels describe different parts of the arrangement.

What happens when AI moves onto your computer

With local AI, the computer doing the model’s work is your own device rather than a remote provider’s server. You download a suitable model and use software that can run it. Tools such as Ollama support this kind of setup. Once the necessary files are present, a fully local workflow can operate without internet access, provided it does not rely on online features. [11]

You are usually running a model somebody has already trained. The technical name is inference: using a trained model to produce an output. This is much less demanding than recreating its original training. Downloading a model is closer to bringing home a finished instrument than manufacturing the instrument from raw materials. [12, 13]

For a person, that might mean help rewriting a letter or explaining notes. For an organisation, it could mean running a suitably tested assistant on its own equipment. These are possible uses, not a claim that any downloaded model will perform them reliably.

The limits are physical. Model size, available memory and the processor or graphics hardware affect what runs and how quickly it responds. A common technique called quantization stores model numbers at lower precision to reduce their size and memory needs. It can make local use more practical, with possible losses in accuracy. The result has to be tested for the intended task. [11, 12]

Local processing can also keep prompts away from a hosted model provider. But the whole application matters: an online search feature, a cloud model option or another connected service can still send information elsewhere. “Installed on my computer” is not, by itself, proof that every part of an AI system stays there. [11]

Nor does local use make computing costless. You supply the hardware, electricity, storage and some responsibility for updates. You may gain control over a particular downloaded version, while taking on work that a hosted provider would otherwise handle.

A further step is fine-tuning: continuing a pretrained model’s training on a smaller set of examples for a particular purpose. It can help adapt behaviour without repeating the original training process. Simply chatting with a local model does not itself perform that training step. Saving a conversation and updating the model’s weights are different operations. [13]

Where frontier companies fit

A frontier model is a model near the leading edge of capability. “Frontier” describes ambition and performance; “open source” describes permissions. They do not divide the industry into two completely separate camps.

With a hosted proprietary service, you access a model through an application or a developer connection, often called an API. The provider operates the model. You generally cannot take its private weights home and run that exact service independently. A downloadable model gives you another deployment option, subject to its licence. [14]

Large companies can participate in both approaches. OpenAI released the open-weight gpt-oss-120b and gpt-oss-20b models under Apache 2.0 on 5 August 2025. Alibaba’s Qwen team also releases downloadable models; the Qwen3 repository identifies its open-weight releases as Apache 2.0. Those examples do not make every product from either company open. [14, 19]

The practical choice depends on the job. Hosted services can spare you the work of operating a model. Local or independently hosted models can offer more control over deployment and changes. Neither arrangement proves that an answer is correct. Capability, reliability and cost need to be assessed for the specific system and task.

Five names that show the range

These five examples are a guide to different parts of the ecosystem, rather than a ranking of the world’s “best” open-source organisations.

Linux

A software project, rather than a single consumer brand or AI company. Its kernel sits at the centre of many complete operating systems. Linux shows how shared software can become infrastructure used across servers, devices and commercial products. You can depend on it without ever deliberately opening an application called Linux. [5]

Mozilla and Firefox

Mozilla is the organisation associated with the Firefox browser. Firefox shows what open source looks like in an everyday application: its source code can be studied and modified under its licensing terms. Mozilla’s names and logos have separate trademark rules. Permission to reuse code does not automatically let someone present a modified browser as an official Mozilla product. [15]

Blender

Blender is a creative software project for work including 3D modelling, animation and rendering. Its code is released under the GNU General Public License. It is a useful reminder that open source is not only invisible infrastructure or programmers’ tools. It can be the application an artist uses to make a character or a film. [16]

Hugging Face

A company providing a major meeting place for AI development. Its Hub hosts model, dataset and application repositories, while its software helps people work with AI. Think of it as a platform and toolkit, not one enormous AI model. [17] Individual uploads have their own licences; appearing there does not make every upload fully open source. [18]

Alibaba and Qwen

Qwen is a model family developed by Alibaba Cloud’s Qwen team. Its downloadable releases illustrate how a large commercial company can supply building blocks other people run and adapt. The Qwen3 project is a concrete Apache 2.0 example. Anyone choosing a particular model should check that release’s own licence rather than assume the entire brand has one permanent set of terms. [19]

What people think and what we can actually measure

There is no single, demonstrated worldwide public verdict on open source in the evidence reviewed for this article. Someone may value privacy, another may want easy installation, and another may never encounter the term. Surveys of developers and organisations can tell us about those groups; they cannot speak for everybody.

One useful example is the 2026 State of Open Source Report, produced by Perforce OpenLogic with the Open Source Initiative and Eclipse Foundation. Its survey gathered more than 700 responses from people using open source professionally, including engineers, architects and organisational leaders. Avoiding dependence on a single supplier, often called vendor lock-in, was cited by 55% as a reason for adoption. [20]

That is a meaningful finding about those respondents, not a claim that 55% of the general public prefers open-source AI. The same report also describes persistent maintenance and security-update problems. People can value the freedom to operate software independently and still find that independence expensive to manage. [20]

Openness creates the opportunity for scrutiny; someone still has to carry it out. An available codebase can contain mistakes. A downloadable AI can produce unreliable answers. In its gpt-oss model card, OpenAI also notes a particular difficulty with distributed weights: after people have downloaded them, the original developer cannot simply revoke those copies or impose new safeguards on them. [21]

The useful question is therefore more specific than whether open source is good or bad. Who maintains this project? What does its licence permit? What has actually been tested? Those questions turn a broad label into information a person can use.

The possibility inside the laptop

Return to that imagined assistant running at home. Its existence would not mean that one laptop had recreated an entire frontier laboratory. Its abilities would still depend on its model, the available hardware and the care taken in building the surrounding application.

What changed is the starting point. An individual can sometimes begin with work that others have already published and build something useful from it, within the permissions granted. A student can examine a tool. A small team can adapt one. An artist can create with software maintained by people far beyond their own studio.

That possibility connects today’s local AI to the older history of shared software. The person at the keyboard has a chance to become a participant in making the tool, as well as someone who uses it.

Sources and reading

Sources checked on 26 October 2026. Counts are snapshots. Survey findings describe their respondents. The five examples are illustrative.

[1] Open Source Initiative — The Open Source Definition The permissions required for software to qualify as open source.

[2] Open Source Initiative — Frequently answered questions Licensing, commercial use, free software and publicly visible code.

[3] Free Software Foundation — History The GNU announcement in September 1983 and the start of work in January 1984.

[4] Open Source Initiative — History of the OSI Earlier sharing traditions, the February 1998 naming discussion and the creation of OSI.

[5] Linux Foundation — It is not just the Linux operating system Linux origins in 1991 and the distinction between a kernel and a complete operating system.

[6] GitHub Open Source Guides — How to contribute Contributions, maintainers, proposed changes and work beyond programming.

[7] Red Hat — What is open source A commercial participant explains community development and paid enterprise support.

[8] GitHub — Octoverse 2025 The 395 million figure is a 2025 count of public repositories, not a worldwide count of distinct open-source projects.

[9] Hugging Face — Model catalogue Live catalogue checked on 26 October 2026. It displayed 3,189,129 models; the article rounds this to over three million.

[10] Open Source Initiative — Open Source AI Definition 1.0 Requirements for parameters, code and training-data information. Full publication of every raw training example is not required.

[11] Ollama — Frequently asked questions Local and cloud operation, data handling, hardware use and disabling cloud features.

[12] Hugging Face — Quantization concepts Reducing numerical precision, memory requirements and possible accuracy tradeoffs.

[13] Hugging Face — Fine tuning Continued training of a pretrained model on a smaller task-specific dataset.

[14] OpenAI — Introducing gpt oss 5 August 2025. The release of the gpt-oss-120b and gpt-oss-20b open-weight models under Apache 2.0.

[15] Mozilla — Firefox notices Open-source licensing and the separate treatment of Mozilla trademarks.

[16] Blender — Official project README The project’s creative functions and GNU General Public License.

[17] Hugging Face — Hub documentation A platform for model, dataset and application repositories, with community and commercial services.

[18] Hugging Face — Repository licences Repository-specific licensing. Publication on the Hub does not establish that every item has an open-source licence.

[19] Alibaba Cloud Qwen team — Qwen3 repository Project authorship, running and adapting the models, and the Apache 2.0 licence statement for its open-weight releases.

[20] Open Source Initiative — The 2026 State of Open Source Report 28 April 2026. Summary of a survey of more than 700 open-source users, conducted with Perforce OpenLogic and the Eclipse Foundation.

[21] OpenAI — gpt oss model card Model limitations and the difficulty of revoking or changing safeguards after downloadable weights have been distributed.