Muse Glimmer 30B

Meta released Muse Glimmer, a 30-billion-parameter multimodal model distilled from Muse Spark, tuned for always-on local agent workflows and shipped under Apache 2.0. To make the 30B parameter model practical for local execution, Meta compresses it to roughly 4-bit and adds block-level speculative decoding so it can an According to the article, the model is aimed at solo developers and startups running it on a 24 GB GPU or an M4/M5 Max Mac, mid-market teams needing on-prem inference without per-token billing, and regulated enterprises that require an air-gappable agent. Meta recommends adding system-level guardrails rather than shipp

Muse Glimmer 30B, developed by Meta Superintelligence Labs, is presented as a serious launch-day option for local agents on 24GB or 32GB machines. The model is a dense causal transformer plus perception encoder with about 29.6B total parameters (including roughly 1.8B for the vision encoder), 52 layers, a 6,656 hidden The article notes the attention pattern uses three 2,048-token sliding-window layers followed by one global layer, repeating as a hybrid design that reduces but does not eliminate long-context memory cost. Input is text and images with text output; video is sampled as frames with no audio input or output, and the knowl