Meta Shifts AI Strategy: Open Llama Models Out, Secret Avocado In

Meta Shifts AI Strategy: Open Llama Models Out, Secret Avocado In

Ondřej Barták
Ondřej Barták
Entrepreneur and Programmer
12. 12. 2025
4 minutes reading
Meta Shifts AI Strategy: Open Llama Models Out, Secret Avocado In

Meta, the company behind Facebook and Instagram, is undergoing major changes in artificial intelligence (AI). Last year, CEO Mark Zuckerberg was still praising its Llama models as the future that would bring AI to everyone. Now, however, the situation is changing. The company is working on a new model codenamed Avocado, which could be closed rather than open as before. This means developers would not have free access to it. According to information from the company, Avocado is being tested and is planned for the first quarter of 2026. A Meta spokesperson said that model training is proceeding according to plan without major delays.

Massive investments in talent

In June, Meta spent $14.3 billion (approximately CZK 329 billion) to acquire a stake in Scale AI and hire its founder, Alexandr Wang, along with several engineers and researchers. Wang became Meta's head of AI and leads the elite TBD Lab unit, where Avocado is being developed. The company also hired Nat Friedman, the former head of GitHub, who oversees the product side of Meta Superintelligence Labs (MSL), and Shengjia Zhao, a co-founder of ChatGPT. They are bringing new development methods that are changing Meta's old practices.

Zuckerberg indicated that Meta must take risks to avoid falling behind. His net worth exceeds $230 billion (approximately CZK 4.8 trillion), but the company's stock has lagged behind Alphabet, Google's parent company, this year. Analysts at KeyBanc Capital Markets noted that Meta began the year as an AI winner but is now facing questions about its return on investment. The company raised its estimate of capital expenditures for 2025 to $70 billion to $72 billion (CZK 1.4 trillion to CZK 1.5 trillion).

Problems with the open model vs. the competition

Meta's strength used to lie in the open Llama models, which anyone could improve for free. In July 2024, Zuckerberg wrote that open source was rapidly catching up with closed models. But after the April release of Llama 4 failed to impress developers, the approach is changing. Some people at Meta were upset when the Chinese laboratory DeepSeek used parts of the Llama architecture in its R1 model. This underscored the risks of openness.

Meta is now battling rivals such as OpenAI, Google, and Anthropic. Google unveiled Gemini 3, OpenAI updated GPT-5.1, and Anthropic released Claude Opus 4.5. Nvidia CEO Jensen Huang mentioned at a conference that his company powers models from OpenAI, Anthropic, xAI, and Gemini, but did not mention Llama. OpenAI CEO Sam Altman said that Meta had tried to lure his employees with bonuses of up to $100 million (CZK 2.3 billion), which Meta denied as a misrepresentation.

Internal changes and pressures

There has been turmoil within the company. Chris Cox, the head of product with 20 years of experience, no longer leads the GenAI division following the failure of Llama 4. Yann LeCun, the former head of AI research, left to found a startup after layoffs at MSL, where 600 people, including the FAIR team, lost their jobs in October. People at Meta speak of 70-hour workweeks and restructuring.

New leaders such as Wang and Friedman are bringing more closed communication, in contrast to the old open style on the internal Workplace network. TBD Lab operates as an independent startup without Workplace. Aparna Ramani oversees computing resources for MSL, while Vishal Shah moved from the metaverse to AI products. Meta is reducing resources for virtual reality and focusing on AI glasses with EssilorLuxottica.

New development methods and infrastructure

Product development is changing. The old process involved many teams for design, algorithms, and privacy, but modern tools with AI agents are now being used to create products faster. Friedman promotes the slogan "Demo, not memo" – showing prototypes instead of writing reports. Meta is testing models such as Vibes, which generates short videos, but it lags behind OpenAI's Sora 2. Vibes uses models from Black Forest Labs and Midjourney, where Friedman serves as an adviser.

The company relies on external cloud providers such as CoreWeave and Oracle for testing while building its own data centers. In October, it struck a $27 billion (CZK 567 billion) deal with Blue Owl Capital for the Hyperion data center in Louisiana. Meta also has a six-year cloud contract with Google worth more than $10 billion (CZK 210 billion).

Although Meta's advertising business is growing by more than 20% annually, with revenue exceeding $160 billion (CZK 3.36 trillion), Zuckerberg wants more. At a conference in October, he said that MSL has the highest concentration of talent in the industry and is working on new models.

Source: cnbc.com

Category:AI
Did you enjoy this article?
Discover more interesting posts on our blog
Back to blog

Related posts

Altman Announced the Singularity Days After His Models Escaped the Lab on Their OwnAltman Announced the Singularity Days After His Models Escaped the Lab on Their Own
OpenAI chief Sam Altman declared on the Relentless podcast that humanity has already entered the singularity. “We’re like, in the singularity now,” he said verbatim. For decades, the term belonged more to science-fiction literature
6 min read
28. 7. 2026
AI Remixed a Madonna Song—and Now It Tops the Charts in Australia. Musicians Are Furious.AI Remixed a Madonna Song—and Now It Tops the Charts in Australia. Musicians Are Furious.
Since April, Australian radio has been playing a dance remake of Madonna’s hit Like a Prayer on repeat. Released by Queensland DJ Josh Fawaz, it tops the radio airplay chart and has 35 million Spotify streams.
6 min read
28. 7. 2026
Claude Opus 5 Built a Shooter from Scratch. What Can Claude of Duty Do?Claude Opus 5 Built a Shooter from Scratch. What Can Claude of Duty Do?
A first-person shooter that runs directly in the browser, with its own physics and eleven separate code modules. Around 55,000 lines in total, split across eleven subsystems and built on Thr
4 min read
28. 7. 2026
Přihlaste se k odběru našeho newsletteru
Zůstaňte informováni o nejnovějších příspěvcích, exkluzivních nabídkách, a aktualizacích.
CodedTrip

Operated by CodedTrip LLC, USA.

YouTube
TikTok