I have always been a fan of Meta AI—or Facebook AI Research, as many of us still remember it—and particularly of Yann LeCun’s work.
That admiration is not based only on Llama. Facebook has one of the strongest open-source records of any major technology company.
It created React, React Native, GraphQL and Jest for web and mobile developers. Its work on infrastructure produced technologies including Cassandra, Presto, RocksDB, Thrift and Zstandard. In AI, it gave the industry PyTorch, FAISS, Detectron and Segment Anything. Facebook even extended that philosophy to data-centre hardware through the Open Compute Project.
Not every project remained dominant, but together they helped build important parts of the modern technology stack. Meta repeatedly turned technology developed internally into infrastructure that the wider industry could use.
Then came Llama.
Meta was not the first organisation to publish model weights, and the Llama models were not fully open source under the strict definition used by the Open Source Initiative. Nevertheless, Llama transformed the market.
For the first time, developers had broad access to genuinely competitive large language models from a major American technology company. Llama encouraged new inference engines, quantisation methods, fine-tuning tools and local AI applications.
Perhaps the best example is llama.cpp.
Created independently by Georgi Gerganov shortly after the original LLaMA appeared in 2023, llama.cpp demonstrated that large language models could run efficiently on ordinary laptops and desktops—not just expensive cloud infrastructure.
It has since become one of the foundations of local AI, supporting models from Meta, Alibaba, Google, Microsoft, Mistral and many others. Its GGUF model format and quantisation tools have helped developers run increasingly capable models on Macs, PCs and even mobile devices.
llama.cpp was not built by Meta, but it illustrates the wider value of Meta’s open-model strategy. By releasing Llama’s weights, Meta gave the community something around which it could build an entirely new generation of edge-AI infrastructure.
That influence spread globally. Today’s leading Chinese models—including Alibaba’s Qwen and DeepSeek—are independently developed systems, not copies of Llama. But Llama helped establish the tools, culture and competitive precedent that accelerated the worldwide open-weight movement.
That is why Meta’s subsequent pause was concerning.
After releasing Llama 4 Scout and Maverick in April 2025, Meta never released the much larger Behemoth model it had previewed. The company reorganised its AI operation around Meta Superintelligence Labs and appeared to be prioritising powerful closed models. Yann LeCun’s departure further reinforced the feeling that one era of Meta AI had ended.
For roughly 16 months after Llama 4, Meta released no major new open-weight foundation model. Meanwhile, Chinese laboratories increasingly became the centre of open-model innovation.
Now Meta is back.
Muse Glimmer is a 30-billion-parameter model designed to power local AI agents. Released under the permissive Apache 2.0 licence, its weights can be downloaded, modified and deployed by developers.
Glimmer can work with text and images, call tools, handle files, write and debug code, and complete multi-step tasks. Crucially, it is optimised to run through llama.cpp on suitably equipped consumer hardware. That could allow personal agents to work with sensitive information without constantly sending it to the cloud.
More importantly, Mark Zuckerberg has made Meta’s renewed direction explicit. In his manifesto, The Future Is for Everyone, he writes:
“Now that Meta Superintelligence Labs are up and running, we will resume releasing some open source models soon.”
The word “some” matters. Meta is unlikely to release every frontier model it develops. Its strategy appears to be hybrid: retain control of certain cutting-edge systems while releasing capable and efficient models for developers and local devices.
But this is still very good news.
The United States needs a strong open-model champion. Competition between American and Chinese open models will create more innovation, experimentation and choice—and prevent the future of AI from being controlled entirely through proprietary cloud APIs.
Meta has the talent, infrastructure, capital and history to lead the American open-model ecosystem again.
After a worrying pause, Meta is back—and the global open-AI community should welcome its return.
Leave a Comment