Meta Platforms is preparing a major step for Llama 3: the largest version of its open-source language model is planned for release on July 23, according to an employee cited in the source article. The model is described as a 405-billion-parameter system, and it is expected to arrive with weights.
That detail matters. For developers, researchers and the open-source AI community, weights are not a minor technical extra. They are the part of a trained AI model that makes it useful in practice.
A larger Llama 3 with multimodal abilities
The planned Llama 3 release is described as Meta Platforms' largest open-source language model so far. It is also expected to be multimodal, meaning it can process both images and text.
According to the source article, The Information reported that the 405-billion-parameter model should be able to generate new images from a combination of images and text, for example. That would mark an important shift from previous Lama models, which were limited to text generation.
The article does not describe every capability of the new model. What it does make clear is the direction: Meta Platforms is moving Llama 3 beyond text-only use and toward a model that can work across more than one kind of input.
For users and developers, that changes the practical scope of the model. A text-only system is useful for writing, summarizing, coding and answering prompts. A multimodal model can connect written instructions with visual material, which can make it more flexible for applications where images and language need to be handled together.
Why releasing the weights is the central issue
The most closely watched part of the planned release is not only the model size. It is whether Meta Platforms makes the weights available.
In AI models, weights are key parameters that are optimized during training. They help the model make predictions and perform tasks. Without them, developers have only the structure of the system, not the learned capability that makes the model work.
The source article describes a model downloaded without weights as an "empty shell". The architecture may define the neural network, including how layers are arranged and connected, but it does not provide meaningful performance on its own.
That is why weights are valuable to the open-source AI community. When weights are published, developers who do not have massive training capabilities can still use and build on advanced AI models. The source article also notes that publishing weights supports reproducibility, practical application, transparency and comparability.
In plain terms, weights can turn a model release from a technical outline into something people can actually run, test, compare and adapt.
The debate around access and risk
The source article says there were rumors that Meta would not make the weights of the 400 billion model available. AI leaker Jimmy Apples reported on X about alleged objections from Facebook co-founder Dustin Moskovitz to Mark Zuckerberg.
Despite those alleged objections, Jimmy Apples said Meta "apparently at the moment of this update" decided to publish the model, including the weights, as open source.
The debate reflects a basic tension in open-source AI. On one side, publishing weights can make advanced models more accessible and easier to examine. On the other side, the source article notes that there are financial and safety reasons against releasing them.
- Financial reasons: model training is expensive, so releasing the trained result with weights gives others access to something costly to produce.
- Safety reasons: with weights, an open-source model is easier for more people to use directly, which can be criticized from a safety perspective.
- Developer access: weights allow people without massive training capabilities to use and develop advanced AI models.
Those points explain why this release is about more than a single model. It is also about how much of a powerful AI system should be made directly available to the wider developer community.
What developers get with a complete release
If the Llama 3 model is released with weights, developers receive more than a blueprint. They receive a pre-trained model that can make meaningful predictions and solve tasks without requiring them to repeat the full training process.
That distinction is especially important because, as the source article explains, training can be time-consuming and resource intensive depending on the size of the model and the amount of data. A 405-billion-parameter model sits at the center of that concern because its scale makes the weights especially significant.
For the open-source AI community, access to weights can make it easier to test claims, compare behavior and explore practical uses. It can also make the model more broadly usable outside the organizations with the resources to train systems at that scale.
At the same time, the same accessibility creates the safety concern noted in the source article. If more people can use a model directly, then the benefits and risks both become easier to distribute.
That is why the planned July 23 release is likely to draw attention. The model combines three important elements from the source article: it is the largest version of Llama 3, it is expected to be multimodal, and Meta Platforms has apparently decided to publish it with weights as open source.