What open weights leave out that open source AI requires

The Open Source Initiative's Open Source AI Definition, version 1.0, says an Open Source AI is an AI system made available under terms and in a way that grant four freedoms.

An Open Source AI is an AI system made available under terms and in a way that grant the freedoms 1 to:

The four freedoms are to use the system for any purpose and without having to ask for permission; to study how the system works and inspect its components; to modify the system for any purpose, including to change its output; and to share the system for others to use, with or without modifications, for any purpose.

Use the system for any purpose and without having to ask for permission. Study how the system works and inspect its components. Modify the system for any purpose, including to change its output. Share the system for others to use with or without modifications, for any purpose.

The definition says a precondition to exercising these freedoms is access to the preferred form to make modifications to the system.

A precondition to exercising these freedoms is to have access to the preferred form to make modifications to the system.

The definition says AI weights are the set of learned parameters that overlay the model architecture to produce an output from a given input.

AI weights are the set of learned parameters that overlay the model architecture to produce an output from a given input.

For a machine-learning system, it says the preferred form to make modifications must include three elements: data information, code and parameters.

The preferred form of making modifications to a machine-learning system must include all the elements below: Data Information ... Code ... Parameters

Data information means sufficiently detailed information about the data used to train the system, so that a skilled person can build a substantially equivalent system.

Data Information: Sufficiently detailed information about the data used to train the system so that a skilled person can build a substantially equivalent system.

Code means the complete source code used to train and run the system, and parameters means the model parameters, such as weights or other configuration settings.

Code: The complete source code used to train and run the system. ... Parameters: The model parameters, such as weights or other configuration settings.

The definition says "Open Source models" and "Open Source weights" must include the data information and code used to derive those parameters.

“Open Source models” and “Open Source weights” must include the data information and code used to derive those parameters.

It also says it does not require a specific legal mechanism for assuring that the model parameters are freely available to all.

The Open Source AI Definition does not require a specific legal mechanism for assuring that the model parameters are freely available to all.

The Gemma Terms of Use say they apply to the Gemma models listed in their appendix, and point to a separate licence for Gemma 4.

The terms below apply to Gemma models listed in the Appendix at bottom of this page. For Gemma 4 terms, see the Gemma 4 license .

Google's Gemma Terms of Use, last modified on 1 April 2026, say you may use, reproduce, modify, distribute, perform or display Gemma only in accordance with the agreement.

You may use, reproduce, modify, Distribute, perform or display any of the Gemma Services only in accordance with the terms of this Agreement

They say you must not use the Gemma services for the restricted uses in the Gemma Prohibited Use Policy, which is incorporated by reference, or in violation of applicable laws and regulations.

You must not use any of the Gemma Services: for the restricted uses set forth in the Gemma Prohibited Use Policy at ai.google.dev/gemma/prohibited_use_policy (" Prohibited Use Policy "), which is hereby incorporated by reference into this Agreement; or in violation of applicable laws and regulations.

They also say Google reserves the right to restrict, remotely or otherwise, usage that it reasonably believes violates the agreement.

To the maximum extent permitted by law, Google reserves the right to restrict (remotely or otherwise) usage of any of the Gemma Services that Google reasonably believes are in violation of this Agreement.

The same page says Gemma 4 has its own licence.

For Gemma 4 terms, see the Gemma 4 license .

The Gemma Terms of Use say the terms apply to the Gemma models listed in the Appendix at the bottom of the page.

The terms below apply to Gemma models listed in the Appendix at bottom of this page.

Google's Open Source Blog said on 2 April 2026 that Gemma 4 models are the first in the Gemmaverse to be released under the OSI-approved Apache 2.0 license.

Gemma 4 models are the first in the Gemmaverse to be released under the OSI-approved Apache 2.0 license.

Google's Open Source Blog post on Gemma 4 says Gemma 4 allows local, private execution that does not rely on cloud-only infrastructure.

Gemma 4 allows for local, private execution that doesn't rely on cloud-only infrastructure.

The Hugging Face model card for Qwen3.5-9B lists its licence as apache-2.0.

License: apache-2.0

Anthropic's self-replication post of 12 October 2026 ran Gemma4-26B-A4B distributed over 15 inexpensive off-brand Android TV boxes, about $55 each, at 0.39 tokens per second.

We run Gemma4-26B-A4B, distributed over 15 inexpensive off-brand Android TV boxes (~$55 each) that we bought and networked together, at 0.39 tokens per second.

The same post says GLM-5.2 was the only model in its set of tested open-weight models capable of full end-to-end replication in its simplified test, assuming its capabilities could be packed into Llama3.1-8B.

Only GLM-5.2 out of our set of tested open-weight models is capable of full end-to-end replication (assuming its capabilities could be packed into Llama3.1-8B).

Anthropic's distillation post of 12 October says the US Center for Advancing Innovation and Standards for Super Intelligence has assessed GLM-5.3 as the most cyber-capable open-weights model released to date.

Early evidence suggests that these methods have reduced the amount of data harvested by distillers by around 90 to 99%.

The Open Source Initiative's page on open weights says "Open Weights refer to the final weights and biases of a trained neural network."

Open Weights refer to the final weights and biases of a trained neural network.

The same page says open weights differ from Open Source AI because they do not include training code, which it describes as "The scripts or frameworks used to create and curate the training dataset."

However, Open Weights differ significantly from Open Source AI because they do not include: Training code - The scripts or frameworks used to create and curate the training dataset.

The same page says open weights do not include the training dataset, which it describes as "The full dataset used for training, when legally possible."

Training dataset - The full dataset used for training, when legally possible.

The same page says open weights do not include comprehensive data transparency, which it describes as "Full details about dataset composition, such as source domains, cleaning methods, or balancing techniques."

As an alternative, when distribution of the training dataset is not legally possible, Comprehensive data transparency - Full details about dataset composition, such as source domains, cleaning methods, or balancing techniques.

The same page says that by withholding these elements, developers "only provide a glimpse into the final state of the model, making it difficult for others to replicate, audit, or deeply understand the training process."

By withholding these critical elements, developers only provide a glimpse into the final state of the model, making it difficult for others to replicate, audit, or deeply understand the training process.