Meta Closes the Door on Llama: Muse Spark Marks the End of Big Tech’s Open-Source Era

The company that taught the industry to ship open weights just shipped a proprietary frontier model instead. With talent leaving and a reported nine-figure AMD infrastructure bet behind it, Meta's pivot says more about the open ecosystem than any license file ever could.

For three years, Llama was the load-bearing wall of open AI. Whenever a startup needed a capable model without renting one from OpenAI or Anthropic, Meta's weights were there, free to download and fine-tune. That era now has a clear endpoint. Meta's new flagship, Muse Spark, built by Meta Superintelligence Labs and announced in April 2026, is not Llama 5. It is something the company has spent years insisting it would never lead with: a closed model.

Meta frames Muse Spark as "the first in a new series" and its "most powerful model yet," the product of a roughly nine-month rebuild of its internal AI stack. It is multimodal by default — it reads images to analyze food and nutrition, scans products, generates working websites and apps, holds interruptible voice conversations, and spins up parallel "subagents" to chew through complex questions. It already powers the Meta AI app and meta.ai, with rollout underway across WhatsApp, Instagram, Messenger, Facebook, and Ray-Ban and Oakley Meta glasses.

What it is not, at launch, is open. Access comes through the consumer apps and a private-preview API for select partners. Meta says future versions may eventually be open-sourced — a notably softer commitment than the "open by default" posture that defined the Llama years.

The retreat nobody wanted to name

The uncomfortable part for the open ecosystem is who is making this move. Meta was not a reluctant participant in open-weights AI; it was the anchor tenant. If even Meta concludes that its frontier capability is too valuable — or too expensive to give away — the economic case for openly releasing the best models gets thinner across the board. The remaining open frontier increasingly leans on a smaller cast: well-funded challengers and, conspicuously, Chinese labs whose releases now define much of the open-weights conversation.

That leaves a familiar split. Truly capable, current-generation weights drift behind APIs, while "open" comes to mean the previous generation, or smaller models tuned for efficiency rather than frontier performance. Useful — but not the same thing as an open frontier.

Talent and compute, pulling the same direction

The strategy shift sits alongside two pressures worth naming carefully. Industry reporting has described a wave of senior researcher departures from Meta's AI organization during its reshuffle into Superintelligence Labs — the exact scale and roster are still being reported and should be treated as *to be confirmed*. Separately, a reported infrastructure agreement with AMD valued at roughly $100 billion has circulated as part of Meta's compute buildout; the precise terms remain *to be confirmed* pending confirmation from the companies.

Read together, the pattern is coherent even if individual figures need verifying: when a model costs this much to train and run, and when the people who built it are in play, giving the weights away looks less like generosity and more like forfeiting an asset. Muse Spark is the result of that arithmetic.

The open ecosystem will survive this. It just lost the company that made it look inevitable.

Sources

• VentureBeat — "Goodbye, Llama: Meta launches new proprietary AI model Muse Spark": https://venturebeat.com/technology/goodbye-llama-meta-launches-new-proprietary-ai-model-muse-spark-first-since
• Meta Newsroom — "Introducing Muse Spark from Meta Superintelligence Labs": https://about.fb.com/news/2026/04/introducing-muse-spark-meta-superintelligence-labs/