Most models people call “open source” today are not actually open source. Llama, DeepSeek, Qwen, and Gemma all get labeled that way in headlines, but under the definition the software industry actually uses, almost all of them are open weight instead.
The difference sounds technical, but it decides whether you can legally rebuild the system, redistribute a modified version, or walk away from the original vendor without starting over.
This article breaks down what each term actually means, how to tell them apart, and where the most talked about 2026 models actually land.
Key Takeaways
|
What Open Weight Actually Means
Open Weight means a developer publishes the trained parameters of a model so anyone can download them, run the model locally, and fine tune it.
It does not mean the developer shares the training code, the raw or processed training data, or a full account of how the model was built.
Open Weight models release the trained model weights for download and fine tuning, but do not disclose all training data, code, or development details.
Think of it as being handed a finished car. You can drive it, repaint it, swap the engine, or take it apart to see how the pieces fit together.
What you do not get is the factory blueprint, the supplier list, or the assembly line instructions that produced it. You can study the output, but not fully reproduce the process that created it.
Llama, DeepSeek’s R1 and V3 lines, Alibaba’s Qwen models, and Google’s Gemma line are the most common examples people point to when they say “open source AI.”
Nearly every model marketed that way today, including Llama, DeepSeek, Qwen, and Gemma, is technically open weight only.
What Open Source Actually Means
Open Source AI means a model can be used for any purpose, studied to understand how it works, modified freely, and shared without restriction, backed by real access to the code and data information needed to make that possible.
The OSI defined open source AI models as those that can be used for any purpose without permission, allow the study of the system’s components and how it works, permit modification for any purpose, and allow the system to be shared for any purpose.
This definition, called OSAID 1.0, was finalized on October 28, 2024, after a multi year, co design process involving workshops around the world.
To satisfy it, a release must include the training and inference source code under an OSI approved license, model weights under terms that allow free use and redistribution, and enough data information about training data sources and processing that a skilled person could substantially recreate the system.
That last requirement is the one almost every “open” AI release fails. Companies rarely publish full training data details, both because the data is often licensed or scraped under legal gray areas and because it represents a competitive advantage they want to protect.
A short list of models has actually passed OSI’s own validation process against this bar: Pythia from EleutherAI, OLMo from AI2, Amber and CrystalCoder from LLM360, and T5 from Google. BLOOM, StarCoder2, and Falcon came close but would need license changes to fully qualify.
Read Also: API-First Development: Why APIs Should Be Your First Priority
Where Popular 2026 Models Actually Fall
Licensing has shifted noticeably through 2026, and it is worth checking a model’s current terms rather than assuming last year’s label still applies.
DeepSeek ships its recent models under the MIT license, and Qwen moved its open line through version 3.6 to Apache 2.0.
OpenAI’s gpt oss models also carry an Apache 2.0 license. Google moved Gemma 4 to Apache 2.0 in April 2026, dropping the custom terms earlier Gemma versions carried.
Meta’s Llama models, by contrast, still ship with only inference code and a license that restricts some commercial use cases, which is why researchers do not classify them as open source even though the weights are public.
Permissive licensing is common across the field now. Open source, in the strict OSAID sense, is still rare.
A model can be generously licensed for downloading and fine tuning while still falling short of the fuller openness the term “open source” implies.
Why the Distinction Matters
The distinction affects four practical decisions a team has to make before adopting any model.
Redistribution rights depend on the license, not the open weight or open source label.
Some open weight licenses, including Meta’s Llama terms, place conditions on redistributing modified versions or using the model above certain usage thresholds.
Vendor lock in risk is lower with true open source models, because a team that has the training code and data information can, in principle, retrain or fork the system independently if the original vendor changes terms or shuts down.
Audit depth differs sharply. With an open weight model, you can test the outputs and inspect the architecture, but you cannot trace a problematic output back to a specific training example or data source. With an open source model, that kind of investigation is possible in principle, even if it takes real effort.
Policy and procurement reviews increasingly ask which category a model falls into.
Some public sector and enterprise procurement processes now specifically ask whether training data information was disclosed, in addition to whether weights are downloadable.
How to Check Which Type a Model Actually is?
Here are four checks to run before you call any model open source in your own documentation or marketing.
Check the weights. Confirm you can actually download the trained parameters and run them without a paid API call. If you cannot, the model is closed, not open weight or open source.
Check the training code. Look for a public repository that includes the actual training pipeline rather than only inference or fine tuning scripts. Most vendors release only the latter.
Check the data information. Look for a data card or technical report describing what the training data was, where it came from, and how it was filtered. Few vendors publish this in the depth OSAID 1.0 requires.
Check the license text, not the marketing page. Read the actual license attached to the weights.
An OSI approved license such as MIT or Apache 2.0 is a strong signal. A custom license with usage caps, field of use restrictions, or redistribution conditions is a signal the release is open weight only.
Conclusion
Open Weight and Open Source describe two different levels of access, and most models marketed with either term today are open weight rather than fully open source.
Open weight gets you a working, downloadable, fine tunable model. Open source, under the Open Source AI Definition, gets you that plus the code and data information needed to understand and rebuild the system from the ground up.
Before you build on a model, read the actual license and check whether the training code and data details were released, rather than relying on how the vendor labels it.
Comments are closed, but trackbacks and pingbacks are open.