Meta releases Muse Glimmer, an open-weight 30B agent model optimized for local deployment

Meta released Muse Glimmer, an open-weight 30B-parameter model under Apache 2.0, positioning it as optimized for long-horizon agent loops, tool use, and local deployment. The model is designed to run entirely on consumer hardware, with Meta explicitly noting quantization to bring the language model under 20GB and a lightweight DFlash drafter for faster on-device generation.

Meta’s announcement was accompanied by a product thread and download links, framing Glimmer as part of a renewed commitment to broadly available personal superintelligence. Mark Zuckerberg and Alexandr Wang were cited in coverage of the release.

Community analysis highlighted architectural details such as similarities to Gemma 4-style hybrid attention plus scale-free QK normalization, larger vision depth, and longer sliding-window attention. Reports also noted that Glimmer was logit-distilled from Muse Spark and trained from the outset on agentic traces, distinguishing it from conventional base-then-post-train releases.

Third-party benchmarking from Artificial Analysis placed Muse Glimmer at 35 on its Intelligence Index, just behind Qwen3.6-27B (38) and around Kimi K2.5 (36), while scoring 44 on the Openness Index. Analysts characterized Glimmer as strong for its size category.

Meta also announced plans to release Muse Spark 1.2 weights 'soon', extending its open-weight strategy beyond Glimmer. The broader context includes Meta’s essay outlining a vision for personal superintelligence accessible to individuals, contrasting with other labs focused on institutional use cases.

Coverage emphasized Glimmer’s ability to run on a single RTX 3090, framing it as a small win for open American models and a step toward always-on local agents.