Meta: Muse Glimmer 30B

Glimmer's own card puts the count at about 29.6B, including a 1.8B vision encodersource, accessed 2026-08-28 — most of a 30B model, but not all of it doing the same job; strip the 1.8B vision encoder out and the text backbone is closer to 28B. Size alone is not what gets it onto one card, though. At full precision a 28B backbone is more than twice a 24 GB budget, and the card is explicit about the step that closes the gap: it compresses the weights to roughly 4-bit precision, shrinking the language model to under 20 GBsource, accessed 2026-08-28, which is what leaves headroom for the KV cache, the perception encoder and the drafter inside a single 24 GB envelope. Wikipedia's summary of the result — that Glimmer runs offline on a single 24 GB consumer GPUsource, accessed 2026-08-28 — is a claim about that quantized build, not a consequence of the parameter count: a multimodal model compressed to leave your account out of the loop entirely.

That local-hardware claim is not just a spec-sheet line — this same row also carries an OpenRouter price, $0.35per million tokensopenrouter-models, last checked 2026-09-14 input against $1.50per million tokensopenrouter-models, last checked 2026-09-14 output over a context window of 131072tokensopenrouter-models, last checked 2026-09-14. The identical weights are usable two ways: rented by the token through that API, or downloaded once and run on a single card for every request after. The licence governing the download is Apache License 2.0source, accessed 2026-08-28, which attaches no field-of-use or user-count condition to the second path — a hobbyist with one consumer GPU gets the same weights a paying API customer is billed for.

Glimmer did not start as its own model. Per its card, Muse Sparksource, accessed 2026-08-28 is where it comes from; Glimmer itself was released 2026-08-10source, accessed 2026-08-28. Only one of those two paths is open on the teacher, though. Muse Spark is rented by the token like any hosted model — meta/muse-spark-1.2 is a row in this same catalog — but it does not come with a download. Meta Superintelligence Labs' own record of its flagship line states not released; Meta's 1.3 announcement (2026-09-02) names 'the Muse Spark open weights release' on its roadmapsource, accessed 2026-09-03. The lab kept the larger model behind the API and shipped only the smaller, distilled one as something you can also just own.