Meta Muse

by Meta

Meta’s 2026 open-weight family and stated Llama replacement — spanning Spark (frontier), Glimmer (30B local), Code, Image and Video.

Series hub. Variants below may earn their own notes as coverage deepens.

Variants

VariantReleasedRole
Muse Spark2026-04-08Frontier model; explicit Llama replacement, “rebuilt from scratch”
Muse Spark 1.22026-08-22Coding model challenging Opus 4.8 and GPT-5.6
Muse Glimmer 30B2026-08-10Open-weights agentic model that runs on local devices
Muse Image / Muse Video2026Generative media variants
Muse Code2026Coding tool built on Muse — see [[../../ai-coding/muse-code

Timeline

DateEvent
2026-04-08Muse Spark launches — Meta’s comeback model, replacing the Llama line
2026-08-10Muse Glimmer 30B open-weights, local/agentic
2026-08-12Benchmarked against Qwen 3.6 27B as a local model
2026-08-22Muse Spark 1.2 targets Opus 4.8 and GPT-5.6 on coding

Notes

The strategic point is that Meta retired the Llama branding: Muse Spark is described as
rebuilt from scratch rather than an incremental Llama successor.

Reception is not uniform. At least one direct August 2026 test of Muse Glimmer against Nemotron
3.5 Lightning rated one of the pair very poorly — worth checking before treating Glimmer’s
benchmark claims as settled.

Glimmer 30B sits on the boundary between open-weight and small/on-device: 30B is local-runnable
on a workstation but not on a phone.