The End of Openness: Meta Replaces Llama With the Closed Muse Spark Model

A sudden social media post from the head of Meta's frontier AI unit has closed an entire chapter in the development of artificial intelligence. Alexandr Wang, skipping big press releases and the usual corporate pomp, announced a radical technological turn: the company is officially winding down support for the open Llama family of language models and replacing it with a proprietary system, Muse Spark. The news hit the global developer community like a bolt from the blue.

It is not just a product-lineup refresh but a fundamental shift in the tech giant's strategy. An industry that had come to see the company as the main champion and driver of open-source neural networks has suddenly run into the hard reality of closed software and a "walled garden" business model.

For the past three years, Meta carefully built a reputation as the leading counterweight to closed commercial systems, offering free access to powerful model architectures. The Llama series became the foundation for thousands of startups, independent research labs, and hobbyists who could not afford the enormous cost of training their own base models. That altruism, even if driven by a pragmatic wish to undermine competitors' monopolies, created a unique ecosystem in which innovation grew out of the collective effort of the global community. Abandoning that approach in favor of Muse Spark wipes out years of trust-building with developers and leaves a large segment of the market without a critical technological foundation.

The switch to Muse Spark was not a spontaneous decision. It was the result of a large-scale internal overhaul that lasted the past nine months. According to Wang, who took over leadership of the company's core AI division, Meta rebuilt its technology stack from scratch.

It created an entirely new compute infrastructure, adopted a different neural network architecture, and set up new data-processing pipelines. The result of that long effort is Muse Spark, which is already integrated into the company's products as the new intelligence core of Meta AI. For the professional community, though, the biggest shock was the system's total opacity: the new model's weights are locked away on the company's servers, and its source code has become a trade secret.

The first independent tests, run by researchers right after the quiet release, paint a decidedly mixed picture of the new architecture. On standard sets of evaluation prompts, it is clear that the developers made a serious bet on versatility of perception. The model delivers impressive results on complex multimodal reasoning tasks and handles mixed data types with ease. At the same time, experts note a noticeable and unexpected lag in writing code, a skill that is among the most in demand with developers. Management's promises to consider publishing open weights "someday" sound far too vague, and the market reads them more as a polite brush-off from corporate lawyers.

For the global tech industry, this reversal brings severe consequences and unprecedented uncertainty. Companies that invested millions of dollars and months of work in building services on top of the Llama architecture have suddenly lost a direct path to upgrading their systems. With access to Meta's next generation of technology now closed, startups face a hard choice: accept gradual stagnation of their current products, or begin a costly and technically complex migration to alternative open solutions from other vendors. This upsets the fragile balance in a market where open models served as a natural check on corporate price-setting.

Killing Llama to save and monopolize its own AI business shows that the era of romantic openness in the corporate sector has reached its logical end. Meta is making it clear: once it comes to monetizing generative models and integrating them deeply into consumer services, the charity stops. Developers around the

*Meta has been designated an extremist organization and banned in Russia.