Meta is developing a powerful new AI model called “Avocado,” expected to launch as Llama 5, which reportedly outperforms previous models in efficiency and capability even before post-training. After past controversies and internal restructuring, Meta is under pressure to restore its reputation and may shift away from open-sourcing its models to better compete in the rapidly evolving AI landscape.
Meta is developing a new AI model, codenamed “Avocado,” which is expected to be released as Llama 5. According to a reputable internal memo, Avocado is Meta’s most capable pre-trained base model to date, outperforming the best open-source base models even before undergoing post-training. This is significant because most AI models only reach their full potential after post-training, which involves refining the model with human feedback to improve safety, reliability, and instruction-following. The fact that Avocado is already competitive with leading post-trained models in knowledge, visual perception, and multilingual performance suggests that Meta’s core AI research is advancing rapidly.
The video highlights Meta’s recent struggles in the AI space, particularly the disastrous release of Llama 4. Meta was accused of manipulating benchmark results to make Llama 4 appear better than it was, leading to a major PR crisis and the resignation of several team members. This scandal prompted Mark Zuckerberg to overhaul Meta’s AI division, sidelining the previous team and launching the new Super Intelligence Labs. The company is now under pressure to restore its reputation and cannot afford another misstep with Llama 5.
One of the most impressive claims about Avocado is its dramatic improvement in compute efficiency. Internal memos state that Avocado achieves ten times the compute efficiency on text-related tasks compared to previous models, and over one hundred times compared to an unreleased version of Llama 4 called Behemoth. These gains are attributed to higher-quality data, investments in model infrastructure, and deterministic training methods that ensure consistent results. Such efficiency improvements are crucial for reducing the energy use and cost of AI development, helping Meta compete with other tech giants.
The video also discusses the competitive landscape of open-source AI, noting that new models are being released at a rapid pace, with each iteration quickly surpassing the last. Meta’s ability to keep up will depend on its willingness to innovate and ship new models quickly, as demonstrated by Elon Musk’s xAI, which has released multiple major versions of its Grok model in less than two years. Meta’s vast resources, including a massive ad business and a rebuilt AI team featuring high-profile hires, position it well to regain a leading role in the AI industry.
Finally, there is speculation that Meta may choose not to open-source Avocado, marking a significant shift in strategy. By keeping the model closed, Meta could better protect its intellectual property and compete more effectively with rivals who have previously benefited from open-source releases. The key lesson from Meta’s past mistakes is the importance of shipping models quickly and iterating publicly. With its renewed focus, resources, and talent, Meta is poised to become a dominant force in AI if it can maintain momentum and avoid repeating previous errors.