MENLO PARK, 31 AUG 2026 — Three weeks ago Mark Zuckerberg published 6,500 words on open AI and promised the weights for Muse Spark 1.2, Meta's frontier model. We checked the repository. It holds four items, all of them Muse Glimmer, and no Spark.
What we looked at
Meta's Hugging Face organisation contains exactly four repositories: Muse Glimmer 30B, its GGUF build, an ExecuTorch package and an assistant variant. All four were created on 9 or 10 August. There is no Muse Spark repository of any version.
That is not evidence of a broken promise. Zuckerberg gave no date beyond soon, and coverage at the time reported the weights would arrive in the coming weeks, which has not run out. It is evidence that, three weeks after the announcement, the thing being celebrated has not happened.
What did ship, and why the licence is the news
Muse Glimmer is real and substantial. It is a dense causal transformer of about 29.6 billion parameters across 52 layers, with a roughly 1.8 billion parameter vision encoder attached, a context window above 131,000 tokens, and interleaved text and image input.
The part that matters most is four words in the licence file. Glimmer is Apache 2.0, which permits unrestricted commercial use, modification and redistribution. Llama never was. It shipped under a bespoke community licence carrying conditions including a cutoff at 700 million monthly active users, which excluded precisely the companies large enough to compete with Meta.
To frame this as a return to open source misses the point. Meta published weights throughout the Llama years. What it did not publish was an open-source licence, and that is what changed.
Read the benchmark table both ways
Meta's own comparison puts Glimmer against Google's Gemma 4-31B and Alibaba's Qwen3.6-27B, and it leads on several agentic tests: MCP Atlas at 75.5, DeepSearch QA at 74.6, SWE-Bench Pro at 51.2.
The same table has Qwen ahead on two. OSWorld-Verified goes to Qwen at 75.6 against Glimmer's 65.9, and TerminalBench 2.1 at 60.7 against 51.7. Both of those measure a model driving an actual computer, which is the capability most often meant by the word agentic.
Two other scores in the set deserve attention for their absolute value rather than their ranking. Reported figures include 23.5 on a banking task suite and 43.3 on GAIA2. A leading score of 23.5 on a 100-point test means the whole field is failing, not that the winner has succeeded. A table read only for rankings hides that.
The derivative ecosystem is the real measure
Meta's own Glimmer 30B repository shows 591,833 downloads. A community rebuild by unsloth shows 976,125, and an LM Studio community build another 411,485.
More people are pulling a repackaged version of this model than Meta's original. That signals real adoption, not a launch-day metric. The quantised and format-converted builds are what actually runs on a laptop, and they exist because the licence allows redistribution.
The catalogue also includes builds labelled abliterated and uncensored, with 87,575 and 53,979 downloads. Those are the model with its refusal behaviour stripped out. Whatever view you take of that, it is a direct consequence of the licence: Apache 2.0 permits modification and redistribution without conditions, so safety tuning becomes advisory the moment the weights are public. That is the trade being made. Open weights are not costless, and it is better to be explicit about that.
Why a frontier open release would be different in kind
Glimmer is a distilled 30-billion-parameter model. Spark 1.2 is the current frontier system, and opening it would be the first time a major American lab published weights at that tier.
For a startup in Singapore or Jakarta, the consequence is concrete. A 30B model runs on hardware a Singapore or Jakarta startup already owns. A frontier model does not. Serving it requires multi-GPU nodes and the expertise to shard across them — a capability problem, not a licensing one. Open frontier weights would mainly benefit organisations that already have clusters, and would change little for a small team that is better served by a model that fits under 20GB after quantisation.
That is the more useful frame than open against closed. The binding constraint on using a large open model in this region is compute, and no licence relaxes it.
What would settle this
One repository appearing would settle it, and the specifics of that repository are what to read when it does.
Which licence it carries, first. Apache 2.0 for a frontier model would be a different order of commitment from the same licence on a distilled one, and Meta has both precedents available to it. Whether it is the current Spark or a superseded version, second, because opening a model that has already been replaced internally is a lagged release rather than an open one. And whether the weights come with the evaluation harness and the training details, or only the tensors.
Until then, this is a pledge with strong intent and nothing to download.
View author profile → · Editorial policy
