| Model | Score | Status | Released | Price in/out | Context | |
|---|---|---|---|---|---|---|
| Muse Glimmer 30B local agents on a single 24 GB GPU | 7.9 | GA | 2026-08-10 | — / — | 131K | Review → |
| Muse Spark frontier reasoning for consumers; builders via API preview | 7.3 | preview | 2026-04-08 | $1.25 / $4.25 | 1.0M | Review → |
Muse Glimmer 30B is Meta's return to open weights: a 29.6B-parameter dense reasoning model from Meta Superintelligence Labs, released 2026-08-10 under Apache 2.0 with no gated access. It is distilled from Muse Spark, reads text and images, and is built for always-on local agents: tool calling over long sequences, recovery from failed tool calls, and coding. Meta's 4-bit build runs on one 24 GB GPU with full 131,072-token context. The buyer's sentence: the most permissive, easiest-to-run open model Meta has shipped, with strong vendor-reported agent scores, but Qwen3.8-27B, published five days earlier at the same size and license, outscores it on every benchmark both companies report.
Full review →