Meta is attempting a comeback in the large language model (LLM) race after stumbling badly with Llama 4, which was widely panned by the AI community for overstated benchmark claims. The company's new model, Muse Spark, released on April 8, 2026, signals a strategic reset under new leadership, but whether Meta can compete with frontier AI labs like Anthropic and OpenAI remains an open question.

The stakes are enormous. Mark Zuckerberg has invested billions assembling a new AI team, including a $14.3 billion investment in the data labeling startup Scale AI to hire its then-28-year-old CEO Alexandr Wang as Meta's chief AI officer. Wang now leads Meta Superintelligence Labs (MSL), the company's focused effort to build world-class AI models.

What Went Wrong With Llama 4?

Meta's previous model release, Llama 4, launched on April 5, 2025, was supposed to be a triumph. The company claimed that Llama 4 Maverick, the mid-sized model in the series, outperformed OpenAI's GPT-4o and Google's Gemini 2.0 Flash across widely accepted benchmarks. The reality was far different.

Independent testing revealed that Llama 4 performed poorly on nearly every benchmark that mattered. Reddit users were brutal in their assessments. "Genuinely astonished how bad it is," one commenter wrote on a post titled "I'm incredibly disappointed with Llama-4." Another called it a "pathetic release from one of the richest corporations on the planet".

The problem was worse than poor performance. Meta had fine-tuned specific models to excel on prominent benchmarks and reported those results, then released different models to the public. Yann LeCun, Meta's then chief AI scientist, later admitted to the Financial Times that the "results were fudged a little bit". The damage to Meta's credibility was severe. Writer Zvi Mowshowitz concluded that Meta belonged in a category of "AI labs whose pronouncements about model capabilities are not to be trusted".

How Did Meta Rebuild Its AI Team?

After the Llama 4 disaster, Meta went silent on LLM releases for an entire year. But Zuckerberg didn't abandon the effort. Instead, he restructured the company's AI operations from the ground up, beginning in June 2025 when he started recruiting aggressively for MSL.

The compensation packages were extraordinary. One 24-year-old researcher was offered $250 million, including $100 million in the first year, according to the New York Times. Engineers received pay packages that "hovered in the mid-tens of millions of dollars". Meta poached several researchers from OpenAI, prompting the latter's chief of research to write an internal memo saying it felt "as if someone has broken into our home and stolen something".

By August 2025, Meta had recruited more than 50 new researchers and started work on a new model codenamed Avocado. The company laid off 600 researchers from older AI units in October, but the new team kept working. By the end of December, Avocado had completed pre-training.

• Recruitment Strategy: Meta invested $14.3 billion to acquire Scale AI and hire its CEO as chief AI officer, signaling a complete organizational overhaul.
• Talent Acquisition: The company offered unprecedented compensation packages, including $250 million to a 24-year-old researcher and mid-tens of millions to experienced engineers.
• Team Consolidation: Meta laid off 600 researchers from older AI units while maintaining the new MSL team, creating a focused group dedicated to frontier AI development.

Why Is Post-Training the Real Challenge?

Muse Spark's release on April 8 received mostly positive reviews, or at least not the relentless criticism that greeted Llama 4. But experts caution that benchmark scores tell only part of the story. The companies producing today's best models, Anthropic and OpenAI, excel at "post-training," the step that gives a model its personality, creativity, resourcefulness, and ethical grounding.

Post-training is where a good model becomes a great one. It's the subtle art of fine-tuning a model's behavior after its initial training, teaching it to be helpful, harmless, and honest in ways that matter to real users. Meta's metrics-obsessed culture may help the company catch up on raw performance, but it may prove to be a poor guide for the kind of innovation needed once models approach the frontier.

"I don't think Meta's new AI team is there yet. And it's not clear if Zuckerberg will be able to build a team with top-tier post-training capabilities, no matter how many billions of dollars he spends on the effort," noted Kai Williams, an AI researcher at Understanding AI.

Kai Williams, AI Researcher at Understanding AI

The question now is whether Muse Spark represents genuine progress or another overstated release. Initial reviews were cautiously optimistic, but the AI community has learned to be skeptical of Meta's claims. The company's reputation will depend not on benchmark scores, but on whether Muse Spark actually works well for researchers, developers, and users in the real world.

Meta's AI journey illustrates a fundamental challenge in the industry: having the resources to build AI models is not the same as having the expertise to build the best ones. Zuckerberg's willingness to spend billions and restructure his organization shows commitment, but whether that commitment will translate into models that compete with Anthropic and OpenAI remains to be seen.