What happened?

During the week of August 10, 2026, Meta made its open-weight AI model Glimmer publicly available. The model allows users to download and run it on their own hardware, a different approach from Muse Spark, the more powerful model the company keeps behind its API. Alongside Glimmer's release, Mark Zuckerberg published a 6,500-word manifesto arguing that AI should be "for everyone."

On TechCrunch's Equity podcast, Kirsten Korosec, Anthony Ha, and Rebecca Bellan discussed Glimmer and Zuckerberg's letter. The hosts pointed out that Zuckerberg's "AI for everyone" messaging carries some contradictions, since Meta's most powerful model, Muse Spark, remains closed.

Why does it matter?

Meta's open model strategy is presented as a stance against AI development being concentrated in the hands of a few large companies. However, the company keeping its most advanced model closed has drawn criticism for a gap between rhetoric and practice. In another TechCrunch piece, Russell Brandom argued that Zuckerberg's manifesto is a sign of "why people don't like AI."

What we know

  • Glimmer is presented as Meta's open-weight, downloadable AI model.
  • Muse Spark is positioned as Meta's more powerful model, kept behind its API.
  • Zuckerberg's letter runs 6,500 words and argues against the centralization of AI.
  • TechCrunch's Equity podcast covered the topic alongside news about the AI industry's energy costs and a $250 million acquisition.

What's next

It remains unclear how the developer community will respond to Meta's Glimmer, and when Muse Spark might become more broadly available. Whether criticism of Zuckerberg's manifesto will affect the company's future open-source strategy remains to be seen.

The model's technical profile

The podcast discussion barely touches the model itself, yet Glimmer's full name is Muse Glimmer 30B and its profile is what makes the strategy legible:

  • 30 billion parameters, a dense architecture — roughly all 29.6 billion activate on every token, not a sparse mixture of experts.
  • Plus a roughly 1.8-billion-parameter ViT-G/14 perception encoder: the model takes interleaved text and images as input and produces text.
  • A 131,072-token context window, support for more than 100 languages, knowledge cutoff of January 4, 2026.
  • Published on Hugging Face on August 10, 2026.

The licence is where the rhetoric becomes concrete

While the consistency of “AI for everyone” is debated, the most concrete data point sits in the licence: Glimmer ships under Apache 2.0, with no constraint like Llama's monthly-active-user cap. This is Meta's first open-weight model since Llama 4, and markedly freer than the licence restrictions of earlier releases.

So the criticism and the counter can both be right at once: Muse Spark, the strongest model, does remain closed, yet the licence on what was opened is looser than before. For anyone wanting to measure the rhetoric, the place to look is not the manifesto's wording but the licence file.

Where it is strong

The model is tuned for agent tasks and its benchmark results point the same way: 75.5 on MCP Atlas, 74.6 on DeepSearch QA, 47.6 on WildClawBench, 43.3 on GAIA2 and 23.5 on τ³-Banking. Those place it in the same band as Qwen 3.6 27B. The low numbers on the last two matter too: parts of the agent benchmark suite remain hard for every model, so being “tuned for agents” does not mean those tasks are solved.