Meta’s next open-source language model, Llama 3, is being positioned as a major step in the company’s generative AI strategy. According to The Information’s sources, the model is scheduled for release in July and is intended to be on par with GPT-4.
The expected launch matters because Llama has become one of Meta’s central bets in AI: a model family that stays aligned with the company’s somewhat open-source strategy while still aiming at the highest tier of commercial AI systems.
A more responsive Llama 3
One of the biggest reported changes is not just model size, but behavior. The new version is said to be more responsive to users when questions are difficult or sensitive.
Instead of blocking tricky questions too quickly, Llama 3 is expected to provide more context. That would mark a shift from the more conservative approach Meta took with Llama 2, where the company wanted to avoid a PR disaster at all costs.
The difference matters because language models often have to judge meaning from context. A word can be harmless in one situation and dangerous in another. The source article gives the example of “kill,” a term that can carry very different meanings depending on how it is used.
Llama 3 is said to be better at making that distinction. If Meta succeeds, the model could answer more useful questions without simply refusing topics that contain ambiguous language.
Safety, tone, and the response problem
Meta also wants to put someone in charge of the model’s tone and safety training. The stated aim is to make the model’s responses more nuanced.
That is a difficult balance. AI systems are often judged not only by whether they are accurate, but also by how they respond when a user asks something controversial, politically sensitive, or potentially harmful.
The source describes two broad camps in AI safety. One is the highly regulated approach, such as OpenAI’s ChatGPT, GPT-4, and Google Advanced, where models follow clear guidelines for polite and politically correct responses.
The other is the more unconstrained open-source movement, along with the contrasting program of X owner Elon Musk. His approach is described as deliberately offering a chatbot that better represents conservative world views, though the source notes that Grok is still criticized by its target audience as too “woke.”
Llama 3 appears to sit in the middle of that debate. Meta wants a model that can be more helpful than Llama 2 while still managing safety risks. At the same time, The Information reports that three key people in the field of AI safety have left Meta this month alone.
How large could Llama 3 be?
The largest version of Llama 3 could have more than 140 billion parameters. That would make it twice as large as Llama 2, whose maximum size is 70 billion parameters.
Size alone does not settle the question of quality. The source notes that parameters are no longer the only and most important measure of the expected output quality of AI models.
Even a Llama 3 model with more than 140 billion parameters would still be far smaller than the original GPT-4 Mixture-of-Experts model with 1.76 trillion parameters. It is not known how many parameters the GPT-4 models currently running ChatGPT have, though the source says it is probably less.
That comparison gives useful scale without deciding the outcome. Llama 3 may be much larger than Llama 2, but Meta’s real goal is performance: the final model is said to reach GPT-4 quality.
Text-only or multimodal remains open
Another unresolved question is whether Llama 3 will be a pure language model again or become multimodal. In this context, multimodal means able to understand and generate images as well as text.
The source says that decision has not yet been made. That leaves an important part of the product direction uncertain.
A text-only Llama 3 would focus the release on language understanding and generation. A multimodal version would push the model into a broader category of AI systems that can handle image and text tasks together.
Fine-tuning has also not yet begun, according to The Information’s sources. That means the model’s final behavior, tone, and safety profile are still part of the work ahead.
Meta’s broader AI push
Llama 2 was released in July 2023, and Llama 3 is now scheduled for July. The timing points to Meta continuing its model development cycle while keeping Llama central to its generative AI plans.
Meta CEO Mark Zuckerberg is described as one of Nvidia’s biggest customers. The company plans to have some 600,000 graphics cards in use for AI training by the end of the year.
The source also says Meta has the long-term goal of developing AGI, just like OpenAI. In parallel, Meta is developing its own AI chip called Artemis.
Taken together, the reported Llama 3 details show a company trying to compete on several fronts at once: model quality, openness, safety behavior, computing capacity, and long-term AI infrastructure.
The most immediate question is whether Llama 3 can deliver on the GPT-4 quality target while being less rigid than Llama 2. If it can, Meta’s next model could become a more capable tool for developers and users who want an open-source language model with stronger performance and more careful context handling.