Can Llama 4.X Save Meta’s AI Ambitions?

Following the release of the Llama 4 series this April, Meta’s progress in AI has faced criticism for underperformance in tasks such as code generation, reasoning and instruction adherence, compared with rivals. According to insiders, Llama 4.X — reportedly codenamed Llama 4.5 — is being developed by a specialised team within Meta Superintelligence Labs (MSL) identified as TBD, with the aim of addressing computational flaws and performance limits in the Llama 4 line.

The plan is for this update to be unveiled publicly before the end of the year, as Meta seeks to restore its standing in the open-source large language model community. Meta launched MSL in June, appointing Alexandr Wang as its head, with a four-pillar structure and a team tasked with training and scaling large models to pursue artificial superintelligence; internal memos have mentioned ambitions for an so‑called omni model.

During Meta’s Q2 FY2025 earnings call, Mark Zuckerberg highlighted parallel work on Llama 4.1 and 4.2 while preparing for the next generation of AI models, alongside aggressive recruitment of researchers from institutions such as Google DeepMind and OpenAI. Yet, less than two months after MSL’s inception, at least eight researchers, engineers and product leads have departed, underscoring the challenges of attracting and retaining a cohesive superintelligence team.

If Llama 4.X delivers tangible improvements in performance and stability before year’s end, it could strengthen Meta’s position in the open-source community and lay a firmer foundation for its broader AI ecosystem.