iTechGuides is reader-supported. When you buy through links on our site, we may earn an affiliate commission. As an Amazon Associate I earn from qualifying purchases. Learn more
An AI model can be “open weight” without being “open source.” “Open weight” describes a release where the trained model parameters are available to download. The Open Source Initiative (OSI) uses “Open Source AI” for a broader standard in its Open Source AI Definition 1.0, which also requires that people can use, study, modify, and share the system, and that the preferred form for modifying it is available. Weights are one part of that picture. Percona CEO Peter Farkas raised the same concern in a September 2026 interview with The Register, arguing that the label matters for buyers and builders evaluating AI tools.
Why the two terms get mixed up
A model card that says “download the weights here” sounds like an open release, and many readers stop there. But a downloadable file of parameters tells you very little about whether you can rebuild the system, check how it was trained, or change it for your own use. The confusion is practical: the label “open source” usually carries expectations about reuse and modification that a weights-only release may not meet.
Not every company, project, or commentator uses these terms the same way. OSI’s definition is its own published framework, not a binding legal standard and not a consensus that the whole industry has adopted. This article uses OSI’s wording because it is the most specific public definition of “Open Source AI” that is widely cited, and it states its criteria in enough detail to check against a release.
What OSI’s Open Source AI Definition 1.0 requires
The definition is published on OSI’s site as The Open Source AI Definition – 1.0. It has two layers: four freedoms that the system must grant, and a set of materials that must be available so those freedoms can actually be exercised.
#1 Best Overall
The four freedoms
- Use: the freedom to use the system for any purpose, without asking permission.
- Study and inspect: the freedom to study how the system works and to inspect its components.
- Modify: the freedom to change the system for any purpose.
- Share: the freedom to share the system, with or without modifications, for any purpose.
To exercise these freedoms, OSI requires that the “preferred form” for making modifications be available. A release that grants use but withholds the preferred form does not meet the definition, even if people can run the model.
What the preferred form includes for machine-learning systems
For machine-learning systems, OSI names three categories of material:
- Data information: enough detail about the training data for a skilled person to build a substantially equivalent system. OSI’s elaboration covers provenance, scope and characteristics, collection and selection, labeling, processing and filtering, and where public or third-party training data can be obtained.
- Code: the complete source code used to train and run the system. OSI’s examples include code for data processing, training, validation, testing, and inference, the model architecture, supporting libraries such as tokenizers, and the relevant training settings.
- Parameters: model parameters such as weights.
The definition does not require one particular legal mechanism for making parameters free. It allows parameters to be free by nature or to be made free through a license or another legal instrument. What matters is whether the terms actually grant the four freedoms.
Free tools Windows power users keep installed
One-click scans. No signup required.
What OSI counts as “the AI model”
OSI describes an AI model as three things together: its architecture, its parameters (including weights), and its inference code. Weights alone are one component. A release that gives you the weight file but not the architecture definition or the code needed to run inference has not handed over the model as OSI defines it.
Practical checks for any AI release
The table below turns the definition into four questions you can answer from a model’s repository, documentation, and license. The third column shows what a partial release typically looks like.
| Check | What to inspect | Typical gap in a weights-only release |
|---|---|---|
| Parameters | Are the model weights or other parameters downloadable? | Often yes; this is the part “open weight” describes. |
| Data information | Is there enough detail on training-data provenance, selection, labeling, processing, filtering, and where comparable data can be obtained? | Often a model card with a high-level description only. |
| Code | Is the code for data processing, training, validation, testing, inference, architecture, and tokenizers available? | Inference code may be published while training code is not. |
| Freedoms and terms | Do the license terms allow use for any purpose, study, modification, and sharing, and are there conditions on any of them? | Use restrictions, acceptable-use limits, or field-of-use limits can fail the freedom to use for any purpose. |
Do not reduce the question to “weights available” or “weights unavailable.” A release can expose weights while withholding the materials needed to study or modify the system in the preferred form. Treat each row as a separate answer, and describe the release using the answers you find.
What Peter Farkas said, and in what context
The Register published its interview with Percona CEO Peter Farkas on September 18, 2026. The conversation focused on databases for AI and agentic workloads. The points below are Farkas’s views as reported in that article. They are his opinions and operational judgments, not conclusions about every AI deployment.
Recommended Free Tools
Vendor lock-in
Asked whether Percona might extend its open-source approach to AI, Farkas said: “Frankly, AI is a vendor lock-in situation.” The concern here is not only about licensing labels. When a system’s data, code, or operating path depends on one provider’s tooling, moving away from it becomes costly, and a weights download does not remove that dependency. The Register’s report presents the lock-in remark as his view, not as a measured finding.
Operational maturity of open-weight agents
The article reports Farkas’s view that there was not yet a mature, enterprise-ready way to run open-weight agents, and that Percona was exploring what it might contribute. Treat that as a judgment about the state of tooling at the time of the interview. It is time-sensitive and may not describe the current state of any particular product.
Where the “not anti-AI” line fits
The same report quotes Farkas as saying “Percona is not anti-AI,” and “Percona is anti-rebranding Postgres or MongoDB as the AI database.” Read together, these statements express wariness toward vendor lock-in and toward marketing labels. The report does not portray him as opposed to AI itself.
Event context
Percona’s community site lists a talk by Farkas, described as CEO of Percona, titled “Standardizing the Future: Navigating the Chaos of ‘Open Source AI’.” The listing describes the talk as covering the OSI framework and the value of standardizing Open Source AI terminology, and it places the presentation on October 7, 2026, within the October 7–9, 2026 conference. The full listing text was not verified for this article, so the listing is cited only to show that Percona has framed this terminology question as a public topic. It is not a source for what was said on stage.
How to describe a release accurately
Use this sequence before attaching the phrase “open source” to any model or agent:
- Locate the license and the terms of use. Confirm whether they permit use for any purpose and whether they allow modification and sharing.
- Confirm that the parameters are downloadable and identify the exact files published.
- Look for data information covering provenance, collection, selection, labeling, and processing, and for any pointers to where comparable training data can be obtained.
- Look for the training, validation, testing, and inference code, plus the architecture definition, tokenizers, and training settings.
- Choose the label that matches what you found. If the weights are downloadable but other materials or freedoms are missing, describe the release as open weight, and state which elements are missing rather than calling it open source.
OSI’s definition is the reference point for the phrase “Open Source AI” in this article. Where a release falls short of it, the more accurate description is a precise list of what is published, under which terms, and what is not.
The Register’s interview with Farkas is at https://assets.theregister.com/2026/09/18/20268/. The Percona Community talk listing is at https://percona.community/talks/2026/2026-10-07-standardizing-the-future-navigating-the-chaos-of-open-source-ai/.
Quick Recap
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
