Meta's Giant Muse Spark Finally Goes Open — With the Old Llama Strings Reattached

Meta's Hugging Face Observatory listing for meta-models/Muse-Spark-Open-816B-A42B appeared in machine.net's tracking data this week as a brand-new entry, with no download history yet to speak of. That absence of a track record is, in this case, the least interesting part of the story. The model is Meta's long-delayed open-weight release of Muse Spark, the closed flagship it spent most of 2026 insisting would stay closed — and the license attached to it quietly puts back a commercial restriction that Meta's own August open-source release had gone out of its way to remove.

The arc here matters. Muse Spark launched in April as Meta's first fully closed, API-only model, ending the Llama line's run as the industry's default open-weight option. In May, Meta's Chief AI Officer Alexandr Wang told an interviewer the model was simply "not suitable for open sourcing." Then, in August, Meta reversed course: it shipped Muse Glimmer, a 30-billion-parameter distillation of Spark built for local use on high-end consumer hardware, released clean under Apache 2.0. Alongside that release, Mark Zuckerberg posted on X that Meta would "soon" release weights for Muse Spark 1.2 itself, calling Meta "a strong supporter of open source." Two months is a strange definition of soon, and no date was ever given in between.

Markedly different from Llama, from a licensing perspective — shared on a true open source licence without any commercial restrictions.

That's how OpenUK CEO Amanda Brock described Muse Glimmer's Apache 2.0 terms back in August, drawing an explicit contrast with the old Llama Community License, which required large commercial deployments above 700 million monthly active users to negotiate a separate agreement with Meta. Brock's point was that Glimmer broke from that pattern entirely — no revenue gate, no carve-out for big API providers. It was, by her account, actually open, with the usual caveat that withheld training data and training code keep any Meta release short of fully open source by stricter definitions.

The strings are back for the model that matters

The Muse-Spark-Open card tells a different licensing story. It ships under what Meta calls the Muse Model License, which carries a Usage Policy and, notably, a Model-as-a-Service clause: large commercial API providers must strike a separate agreement with Meta before serving the model at scale. That is functionally the same shape as the 700M-MAU threshold Brock praised Glimmer for avoiding — reattached, two months later, to the model that actually matters commercially. Meta gave away the 30B distillation with no strings and gated the 816B original behind exactly the kind of clause that made the Llama license controversial in the first place.

There's a coherent business logic underneath the apparent inconsistency, even if Meta hasn't stated it outright. A 30B model tuned for running locally on a single high-end Mac or PC is a developer-goodwill play; it was never going to be the thing hyperscale competitors route production traffic through. An 816-billion-parameter, 42B-active MoE model is a different asset entirely — exactly the scale at which a large commercial API provider might try to resell Meta's own frontier-class weights back to Meta's potential customers. Giving away the small model costs little and buys back developer trust after the April closure; licensing the big one the way Llama 405B was licensed protects the one thing Meta can still extract value from. Read that way, the "reversal" wasn't really a reversal of philosophy — it was always going to apply selectively, and Glimmer's unrestricted license was the exception, not the new rule.

What the weights actually reveal

Setting licensing aside, the model card discloses architecture details Meta never published while Spark was closed and simply listed as "Undisclosed" in third-party tracking. Muse Spark Open is a mixture-of-experts transformer with 816 billion total parameters and 42 billion active per token, built from 256 experts with 8 routed per token, 62 transformer layers at an 8192 hidden dimension, and grouped-query attention with 128 query heads against 16 key-value heads. A separate 1.8-billion-parameter, 50-layer perception encoder handles vision input. It's the first confirmation of how Meta actually built its flagship post-Llama model, after six months of speculation based on benchmark scores alone.

Those benchmark scores are a second awkward detail for Meta. On Artificial Analysis's Intelligence Index, Muse Spark Open scores 39 — behind both GLM-5.3 (45) and Kimi K3 (44), two open-weight competitors that ship without anything close to an 816B parameter count. The pattern holds on most of the card's own agentic and coding numbers: Kimi K3 edges it on OSWorld Verified and Toolathlon Verified, and both Kimi K3 and GLM-5.3 beat it on DeepSWE 1.1 and Terminal-Bench 2.1. Muse Spark Open does lead on GPQA Diamond-adjacent reasoning territory in places, but the overall picture is a very large model that doesn't convert its size into a clear open-weight lead — which makes the commercial licensing gate look less like protecting a crown jewel and more like Meta hedging on a model it isn't fully confident is the best thing on the market.

One caveat on adoption: Ollama's own tracker added a pull listing for muse-spark-open on day one, but it currently shows zero pulls — far too new to say anything about how developers are actually responding. The interesting signal this week isn't usage, it's the license text itself, and what it says about how selectively Meta is willing to apply the "open again" story it told in August.