Meta releases Muse Spark 1.2 and resets its AI strategy with an open-source manifesto

On August 10, 2026, Meta unveiled two new AI models, Muse Spark 1.2 and Muse Glimmer, and opened the weights of its flagship Muse Spark to the public. CEO Mark Zuckerberg simultaneously published a lengthy manifesto in which he acknowledged that the company trails its competitors and declared open source the only sensible path to leadership in the AI race.

What Meta released and how it works

Muse Spark 1.2 is the company's most powerful open-weights model: it can be downloaded, its architecture can be studied, and it can be modified without the restrictions typical of closed systems from OpenAI or Anthropic. Alongside it, Meta introduced Muse Glimmer, a compact model under a permissive open-source license that runs directly on a laptop without cloud infrastructure. Glimmer was trained using distillation: the smaller model absorbs knowledge from the larger Muse Spark.

Key details of the announcement:

• Launch date: August 10, 2026
• Muse Spark 1.2: flagship model with open weights, available for download and modification
• Muse Glimmer: lightweight version with a permissive license, runs on a personal computer
• Glimmer was created through distillation from Muse Spark
• Meta Superintelligence Labs (founded in 2025) is staffed by Alexandr Wang, former CEO of Scale AI

Why Meta is behind its competitors

Meta's recent models have fallen short of systems from OpenAI, Anthropic and Google on standard benchmarks, and the market has treated Meta AI as a product playing catch-up. Despite founding Superintelligence Labs in 2025 and hiring Alexandr Wang away from his post as CEO of Scale AI, the gap with the leaders has not closed.

At the same time, pressure is coming from the other direction: Chinese open models from DeepSeek and Moonshot are closing in on American frontier systems. According to Anthropic, DeepSeek used distillation to extract capabilities from American models, the very same method Meta uses to train Muse Glimmer. Meta's capital expenditures in 2026 are estimated at $145 billion, a significant share of which goes to Superintelligence Labs projects.

Zuckerberg's manifesto: open source as policy

Opening the weights is only part of the announcement. Zuckerberg published a lengthy text laying out both the business and the political logic of the open-source approach. His main argument: concentrating control over AI will "inevitably lead to outcomes that are unfavorable for everyone else," while most large labs are "focused on building AI for companies, governments or other institutions."

"Our goal is for

American open-source models to be the best in the world," — Mark Zuckerberg, CEO of Meta, from the AI strategy manifesto.

The manifesto includes concrete policy recommendations: protect the practice of distillation from legislative restrictions, expand energy infrastructure, and do not restrict access to foreign open-source models.

"I don't think restricting access to foreign open-source models is effective.

That's how the world works, and the US won't be able to lead if it restricts itself," — Mark Zuckerberg.

What this means

Meta is betting on openness as a strategic weapon: to attract developers, build an ecosystem around its models, and challenge Chinese open source at the same time. Whether Muse Spark 1.2 can actually close the gap with GPT-5.5 and Claude will become clear from independent benchmarks in the coming weeks.

*Meta is designated an extremist organization and banned in Russia.