Meta unveils a new AI: it runs on your PC, and it's free

For years, Meta played the role of the good student of open AI: Llama models could be downloaded for free while rival laboratories kept theirs under lock and key. Then Llama 4 was savaged in spring 2025, research was reorganized around Meta Superintelligence Labs, and the new Muse family arrived in April, entirely closed off.Mark Zuckerberg has just reversed course by releasing the weights of Muse Glimmer, a 30-billion-parameter model, under the Apache 2.0 license.

A 30-billion-parameter agent that runs on your machine

Glimmer is not a flagship model and does not claim to be. It was distilled from Muse Spark, compressed to under 20 GB through quantization, and designed to run on a Mac or PC equipped with a single consumer graphics card. Its specialty is agentic capability in the most down-to-earth sense: managing a calendar, organizing files, calling tools, chaining multiple steps, and recovering when one of them goes wrong. The weights have been posted on Hugging Face along with the documentation, and optimized integrations for llama.cpp, MLX, and ExecuTorch are expected in the coming days.

The license deserves attention because it changes the nature of the announcement. Llama models have never been distributed under a recognized open-source license: the proprietary agreement specifically prohibited companies with more than 700 million monthly users from using them without express permission. The Apache 2.0 license wipes all of that away, commercial use included, and puts Glimmer on the same legal footing as the Chinese models that have been flooding Hugging Face for the past two years. For a European company that was hesitant to build its product on Llama, the distinction is anything but cosmetic.

Why Meta is positioning itself as a champion of American open source

The strategic calculation comes down to a market map.Google releases Gemma while keeping Gemini closed, OpenAI has released a handful of open-weight models without ever touching its flagship models, and Anthropic has released none. By contrast, DeepSeek, Qwen and their peers are rolling out open-weight, competitive models at a steady pace. Meta therefore occupies a spot left vacant among the American giants, and Zuckerberg makes no secret of it: he says he wants to “distribute widely” intelligence rather than concentrate it in a few hands.

Others are targeting the same niche with vastly greater resources. Mira Murati’s laboratory, Thinking Machines, released Inkling in July and then Inkling-Small in early August, two open-weight models, the second of which is also licensed under Apache 2.0. When a fledgling has to choose its battles, Meta is announcing up to $145 billion in investment for the current year alone. The title of champion of American open-source AI is earned above all by dint of data centers.

Making weights open also has very concrete effects on this side of the Atlantic. The AI Act eases the documentation obligations of providers of general-purpose models whose weights and architecture are public. The exemption extends beyond the systemic-risk threshold, set at 10²⁵ floating-point operations. A 30-billion-parameter model obtained through distillation comfortably falls below the threshold. Muse Spark 1.2, the large model that Meta promises to open “soon,” falls into a different category.

The famous Spark 1.2 was indeed released on August 5 with paid access via an application programming interface, and no date has been announced for the promised opening. The market has already ruled: Meta’s stock was up 2.1% before Monday’s opening, after declining by around 10% since January. A 6,500-word test and a downloadable model make for a good day on the stock market.

👉🏻 Follow the latest tech news in real time: add 01net to your sources on Google, and subscribe to our WhatsApp channel.

Source: Meta