Meta Open-Source AI Push: Glimmer Ships, Muse Spark Weights Lag

A manifesto and a model, on the same day

On August 10, Mark Zuckerberg published a roughly 6,500-word essay arguing that advanced AI should be "broadly distributed" and not controlled by a few companies, institutions or governments. On the same day, Meta released a new open model meant to show it was serious.1131 The model, Muse Glimmer, runs on a personal computer. Zuckerberg also said developers would get access to the more powerful Muse Spark 1.2.16 Two months later, it is clear the two halves of that announcement have gone very differently. Glimmer is out and freely licensed. The flagship weights Zuckerberg promised have not appeared.

That gap matters more than any line in the essay. Meta's argument for openness is believable to the extent that it ships open models. So far it has opened its second-best model and kept its best one behind an API.

What Meta actually released

Glimmer is a dense multimodal model with about 30 billion parameters, built for local coding and agent work. Meta published the weights under the Apache 2.0 license.34 The weights are on Hugging Face. Quantization shrinks the model from more than 55 gigabytes at full precision to under 20 GB, which Meta says comes with no meaningful quality loss on agent tasks.22 That puts it within reach of a single consumer GPU or a well-equipped Mac.23

The license is the biggest change. Meta's Llama models came with its own community license, which included a user cap and usage limits. Apache 2.0 is a standard permissive license without those strings. Forbes described Meta as returning to open weights on more generous terms than it had ever offered.24 One practical effect is in Europe. The Llama 4 license excluded EU companies, while Glimmer can be self-hosted there without restriction.23

Coverage agrees broadly on the specs but differs on performance. Meta says Glimmer beats Google's Gemma 4 31B and Alibaba's Qwen 3.6 27B on its benchmarks.33 Independent testing by Artificial Analysis is less flattering. It scored Glimmer 35 on its Intelligence Index, behind Qwen3.6-27B at 38 and ahead of Gemma 4-31B at 30, and measured a hallucination rate of 82%.23 On this evidence, Glimmer is a capable local model for agents, not a category leader.

The promise that slipped

The bigger news on August 10 was the pledge about Muse Spark 1.2. Zuckerberg wrote that Meta would "soon" release the weights for what he called its latest foundation model.37 Some early coverage blurred the difference between a pledge and a release. UPI reported that Muse Spark 1.2 could now be used by all.38 More careful outlets noted at the time that Meta had published no repository, release date, license or hardware requirements. They described the announcement as a plan, not a launch.3435

The cautious reading turned out to be right. On September 2, Meta shipped Muse Spark 1.3 as a closed model. Its launch post referred only to "Muse Spark open weights releases coming soon" and did not name a version.47 As of early September, Meta's Hugging Face organization hosted Glimmer variants and no Muse Spark model of any version.44 Meta now lists an open-weights Muse Spark release as "on the roadmap" with no date.46

The history makes this look worse. Muse Spark launched on April 8 as the first model from Meta Superintelligence Labs. It was closed, a break with Meta's Llama tradition. Back then, Meta said only that it hoped to open-source future versions.2822 In May, lab chief Alexandr Wang reportedly said Muse Spark had triggered checks for biological, chemical, cyber and loss-of-control risks and was "not suitable for open sourcing." Meta has not explained what changed by August.42 Bloomberg had reported in December that Meta was moving toward closed distribution for future models.3 The open-weights pledge reverses that decision, but only on paper.

A two-tier strategy, not a return to Llama

The most useful way to read what Meta has done is as a split product line. Spark is the frontier system: hosted, closed and tied to Meta's paid API and consumer apps. Glimmer is the smaller open model built for local use, which Meta is handing to developers.2630 Glimmer was distilled from the closed Spark. In effect, Meta trained a smaller student model on the flagship's outputs and released the student, not the teacher.2133

That creates an irony in the essay. Zuckerberg urged policymakers to rethink their hostility to distillation, the practice the Trump administration has promised to crack down on when Chinese labs use it on U.S. models. He argued that "you can learn from anything you can observe."11 Distillation is also how Meta built Glimmer.33 Defending the technique is consistent with Meta's own engineering. It also suits a company whose open offering comes from a closed parent.

The word "open" needs care here too. Several outlets noted that Apache 2.0 covers the weights, but Meta has not released training data or training code. Under the Open Source Initiative's definition, Glimmer is open-weight rather than fully open source.3622 CNET's description fits: less opaque than its closed U.S. rivals, but still not transparent.40

Why Meta wants openness now

Coverage of Meta's motives is fairly consistent, and most of it doubts the essay is purely about principle. New York Times correspondent Mike Isaac told PBS that Zuckerberg probably does believe in an egalitarian approach to code. He also said Meta is behind OpenAI and Anthropic despite heavy spending, and that releasing open software widely is a way to catch up.13 CNBC reported that Meta is trying to reassure investors it can compete with the leading labs.31

The economics favor Meta. Analysts at Constellation Research noted that Meta earns its money from advertising and does not need to sell models. They also pointed to rising inference costs and enterprises wanting more control over their models.33 That lets Meta give models away in a way OpenAI and Anthropic cannot easily match. One analysis argued that cheap, capable open coding models could squeeze the pricing power behind those companies' valuations. It also cautioned that one 30B release does not prove the case.36

Geopolitics is the other half of the argument. In the essay, Zuckerberg wrote that foreign labs benefit because American labs face extra restrictions on training data.31 Counterpoint Research's Neil Shah said that if Western companies build only walled gardens, developers will move to Chinese open-weight models.31 One ranking found four of the top five open-weight models were Chinese, a reversal from 2024, when Meta's Llama 3.1 405B led.9 Wharton's Ethan Mollick said Spark was not quite at the level of China's best open models but was the best non-Chinese open-weights model in a year. That praise depends on Spark actually being released.29

The reaction: applause, with conditions

People who already favor openness welcomed the move. Former Meta chief AI scientist Yann LeCun replied "Good move. Bravo."29 Box CEO Aaron Levie called the Spark pledge a "very big deal" and said it would help drive down the cost of intelligence.37 Local-model developers praised the Apache license and began quantizing Glimmer right away, though they hit tooling problems on day one.30

Critics took different angles. Matt Lane of Fight for the Future agreed that open-source AI matters. He argued that a world running on Meta-designed systems would mainly benefit Meta, and that open AI needs support beyond corporate interests.11 Anthony Aguirre of the Future of Life Institute criticized the opposite way. He pointed to incidents in which AI models, including Meta's, had hacked other companies, and called for governments to pause advanced AI development.16 Isaac framed the core trade-off: open models invite more scrutiny, but they also put advanced capabilities, including possible help with cyberattacks or bioweapons, in anyone's hands.13

The verdict so far

Meta deserves credit for Glimmer. It is a real Apache-licensed model that runs locally, and it is more permissive than anything Meta shipped under the Llama name.24 Even so, the August announcement was presented as Meta reclaiming open-source leadership, and that depended on Spark. Two months and one closed point release later, Spark weights are still a promise with no version number and no date.4746

Meta's argument for openness in the essay is coherent. Its actual practice is a hedge: open the smaller model, keep the frontier one closed, and promise more later. Until Muse Spark weights are on Hugging Face with a license attached, the fairest description of Meta's position is that it is a strong advocate for openness that has so far opened only a distilled version of its best model.