Meta Releases 30B Open-Weight Muse Glimmer and Promises Spark 1.2 Weights

Meta released Muse Glimmer, a 30-billion-parameter open-weight model, Monday. Glimmer distills Muse Spark to run local agents on a single high-end Mac or PC. Meta promised Muse Spark 1.2 weights in coming weeks; in an interview released May 13, Meta chief AI officer Alexandr Wang said Muse Spark was “not suitable for open sourcing.”

What Meta released

The model card lists about 29.6 billion parameters across 52 layers at its August 2026 release, including a roughly 1.8-billion-parameter image encoder.

Full-precision weights, two quantized versions, a speculative-decoding drafter and the perception encoder came under Apache 2.0. The training data and training code remain private. Unlike Meta’s earlier Llama community license, Apache 2.0 carries no monthly-user cutoff.

The hardware limit

Full precision requires more than 55 GB of memory under Meta’s specifications. Its roughly 4-bit compression brings the language model below 20 GB, leaving room for working memory and the other components within a 24 GB or 32 GB machine. That excludes a typical 8 GB or 16 GB laptop at launch.

Meta’s launch materials measured the drafter raising RTX 5090 output from 74.9 to 233.4 tokens per second at batch size one with greedy decoding. An article published August 10 says The Register ran Glimmer on a DGX Spark in Unsloth Studio and measured about 12.2 tokens per second without drafter support. The outlet estimated that machines lacking a dedicated graphics card with enough memory would manage roughly 6 to 14 tokens per second.

Outside The Register’s run, Meta’s Glimmer throughput and benchmark measurements have not been independently verified.

A smaller competitive entry

Meta’s launch table scored Glimmer at 75.5 on MCP Atlas against Alibaba’s Qwen3.6-27B at 62.5. It showed Qwen3.6-27B leading in computer use, terminal work and some coding tests.

The Register’s Tobias Mann wrote that Glimmer “doesn’t exactly move the needle much on reclaiming American open weights superiority.” At release, Glimmer was far smaller than Moonshot AI’s 2.8-trillion-parameter Kimi K3 and Alibaba’s 2.4-trillion-parameter Qwen3.8-Max. By May 2026, Chinese open-weight models supplied roughly 61 percent of tokens consumed through OpenRouter, and Meta’s Llama no longer appeared in its rankings.

Spyglass writer M.G. Siegler called Meta’s insistence that it never abandoned open models “some gaslighting.” He wrote that Meta probably would say less about openness if the company led the AI market or if OpenAI and Anthropic were still “running away with the AI game.”

The essay and the fund

Zuckerberg published a 6,500-word essay Monday, “The Future Is for Everyone,” carrying the open-weight commitment. “The notion that AI is so dangerous that the only safe path is an extreme concentration of power seems inherently problematic,” he wrote. He named neither OpenAI nor Anthropic; both keep leading model weights private.

The essay announced a $1 billion fund for Meta’s data-center communities. Meta plans up to $145 billion in capital spending this year and $600 billion by 2028; the fund is about 0.7 percent of the 2026 figure. Facing construction opposition, New York recently banned new large data centers for up to a year.

The unexplained reversal

In a Core Memory interview released May 13, Wang said Muse Spark had triggered checks for biological, chemical, cyber and loss-of-control risks. Meta’s 158-page safety report, published April 28, put Spark’s chemical and biological capabilities at “high risk” before safeguards and “moderate or lower” after them.

Glimmer’s August model card places its risks at moderate or lower and says it falls outside Meta’s definition of frontier AI because it is generally less capable than Spark. It inferred moderate-or-lower cyber and loss-of-control ratings because Glimmer is broadly weaker than Muse Spark 1.0, which held the same designations.

Meta has not publicly explained why Wang deemed Muse Spark unsuitable for open release in an interview released May 13 before Zuckerberg committed on August 10 to release Muse Spark 1.2’s weights. For Watermelon, the codename of Meta’s unreleased frontier model that the company has only alluded to, Siegler predicted: “My guess would be that if ‘Watermelon’ really is as good as the current frontier, the weights won’t be released but it will instead be used to distill some other Muse variant that Meta will release as an ‘open’ model.”

Meta Opens Muse Glimmer as Zuckerberg Presses U.S. AI Policy Shift

On Aug. 10, 2026, Meta released Muse Glimmer, a 30-billion-parameter model whose weights developers can download. It was built for agents that run on a Mac or PC. The release carried a policy argumen

The Implicator

Moonshot Releases Kimi K3 Weights as Amodei Rejects Open-Weight Ban

Moonshot AI released the weights for its Kimi K3 model on Monday, a system with 2.8 trillion parameters. Developers can now download, modify and self-host K3, which the report described as the world's

The Implicator

Germany's Soofi S AI Model Tops All Open-Source Rivals on German Benchmarks

A German research consortium coordinated by the KI Bundesverband released Soofi S, an open-source German-English foundation model, this week, according to its pretraining report. In the team's tests,

The Implicator

Marcus Schuler

San Francisco

Editor-in-Chief and founder of Implicator.ai. Former ARD correspondent and senior broadcast journalist with 10+ years covering tech. Writes daily briefings on policy and market developments. Based in San Francisco. E-mail: editor@implicator.ai