Meta Releases Muse Glimmer, First Open-Weight Model Since Llama 4, Paired With Zuckerberg Manifesto on Distillation

Meta has released Muse Glimmer, a 30-billion-parameter model with weights published under an Apache 2.0 license on Hugging Face. It's the company's first open-weight model release since Llama 4 in spring 2025, ending a stretch of more than a year during which Meta stayed out of the open-model race.

Glimmer is designed for AI agents that run locally on a single consumer GPU, handling tasks like calendar management, file access, and messaging without sending data to the cloud. At full precision the model requires more than 55 GB of memory; Meta quantizes it to roughly 4 bits, bringing it under 20 GB, including image processing. Meta says a smaller helper model accelerates text generation by up to 3.1x, though this figure comes from the company's own testing.

Glimmer was built by distilling Meta's larger Muse Spark model — training the smaller model to replicate the outputs of the bigger one. According to the Wall Street Journal, Meta plans to release an open-weight version of Muse Spark 1.2, currently its strongest model, within weeks.

Benchmark claims, with caveats

Meta compares Glimmer against Google's Gemma4-31B and Alibaba's Qwen3.6-27B, the leading open models in its size class. According to Meta, Glimmer wins most benchmarks, particularly on agentic tasks such as tool use, web search, and long-context handling. Qwen reportedly performs better on desktop-control and terminal tasks; the three models are roughly even on multimodal benchmarks. Meta gathered most of this comparison data itself and acknowledges in its methodology report that its test setup was not tuned for the competing models — a disclosure that warrants treating the results as directional rather than independently verified.

Zuckerberg defends distillation

Alongside the release, Mark Zuckerberg published an essay, "The Future is for Everyone," arguing that superintelligence should be distributed widely rather than concentrated among a handful of labs. He writes that a single benevolent superintelligence "can't exist" and that safety comes from balance among many players.

Zuckerberg explicitly defends distilling other companies' models, writing that the principle worth protecting is "that you can learn from anything you can observe." This puts Meta at odds with OpenAI and Anthropic, both of which have accused Chinese labs of distilling their models without permission. Anthropic CEO Dario Amodei has separately called for tighter export controls on China and warned about frontier-capable open models circulating freely. Notably, The Information reported in June that Meta had restricted its own engineers' use of Claude Code and Codex to prevent their outputs from leaking into Meta's training data — a policy that sits uneasily next to Zuckerberg's public stance.

The compute-auction hint

Buried in the essay is Zuckerberg's clearest signal yet on monetization: free model access for billions of users, with a "dynamic auction mechanism" determining who pays for additional compute, echoing Meta's ad-auction business. No product, timeline, or target customer (consumer, developer, or enterprise) has been specified.

The stakes are considerable. Meta plans up to $145 billion in infrastructure spending this year and $600 billion through 2028, according to the WSJ, without a comparable revenue stream. Zuckerberg floated a data-center-monetization idea on Meta's July earnings call that triggered a stock selloff, and he has acknowledged internal weaknesses in the AI reorganization.

What this means

Meta cannot currently match OpenAI or Anthropic on frontier model capability, so it is repositioning distribution and infrastructure as its competitive edge — a bet that giving away strong open models while selling scarce compute access could become a business model. That strategy depends on demand for Meta's compute materializing at auction-worthy prices, which remains unproven. The distillation debate also exposes a contradiction: Meta argues learning from observable outputs is fair game while restricting its own engineers from doing exactly that with rivals' tools.