Meta was reportedly planning a new version of its Llama model that could approach GPT-4’s capabilities while remaining available for broad use. The project would pair upgraded AI training infrastructure with the company’s open-source strategy, positioning Llama 3 as a potential alternative for developers and businesses.
A more capable model was planned for 2024
The Wall Street Journal reported that Meta intended to develop Llama 3 for 2024. Training was expected to begin in early 2024, after the company upgraded data centers for AI training with H100s.
People familiar with the project told the newspaper that the resulting model was expected to be “roughly” on par with GPT-4 and significantly better than Llama 2. That description sets an ambitious target, but it does not establish how the model would perform across particular tasks or applications.
The reported goal is notable because Llama 2 reaches the level of GPT-3.5 in some applications. It has also been improved by the open-source community through fine-tuning and additional applications. Moving toward GPT-4-level capability would therefore mark a substantial step in the series, even if the model were not intended to surpass OpenAI’s system.
Meta’s open approach comes with limits
Meta CEO Mark Zuckerberg was reportedly committed to keeping Llama models available for free in many use cases, including commercial ones. The source also says the access would not cover every use case, so “open source” in this context does not mean unrestricted use.
That strategy could make a more capable model useful to a wider range of developers and organizations. Users could build on it through fine-tuning and other applications, as they have with Llama 2. The source does not specify the terms or conditions for Llama 3, so the exact boundaries of access remain unclear.
There is also a practical question behind the performance ambition. The source describes GPT-4 as a complex system built from 16 expert networks with about 111 billion parameters each. It suggests that distributing a model with that complexity could be much more difficult, which may help explain why Meta’s reported aim was to come roughly alongside GPT-4 rather than exceed it.
A competitive move in a crowded AI race
The planned release was described as part of Zuckerberg’s effort to reposition Meta as a leading AI company. A Llama model near GPT-4’s level could strengthen that effort by giving developers another high-capability system to work with, especially if it remained free for many applications.
But Meta would not be the only company advancing its models. Google was expected to have released Gemini by then, while OpenAI was also expected to introduce further innovations the following year, though probably not GPT-5. The competition would depend not only on benchmark-level capability, but also on what users could access and build with each system.
Earlier comments attributed to OpenAI engineer Jason Wei had already fueled expectations around Llama 3. In late August, he reported meeting Meta AI developers who said they had enough computing power to train Llama 3 and 4. The Wall Street Journal’s report aligned with those remarks, but the plans and performance claims described in both accounts were still reported expectations.
What the Llama 3 target would mean
If Meta achieved the reported goal, Llama 3 could bring GPT-4-level performance closer to users who want to adapt a model for their own applications. The open approach would offer a different route from relying solely on a proprietary model, while the stated limits on use would still matter to organizations considering adoption.
The comparison with GPT-4 also highlights the challenge. The source presents Llama 2’s community-driven improvements as a foundation, but reaching a more advanced system’s level would be harder. Meta’s infrastructure upgrades, training plans and distribution choices would all shape whether Llama 3 could meet the target.
At the time of the report, the key points remained plans: training was expected to start in early 2024, and the model was projected to be roughly on par with GPT-4. Its eventual capabilities and the terms under which people could use it would determine how consequential the release became.