Meta open-sources Muse Glimmer, a 30B agentic model that runs on a single consumer GPU

Meta released open weights for Muse Glimmer, a 30-billion-parameter agentic model built to run entirely on a single consumer GPU, making it the company's first Meta Superintelligence Labs model aimed specifically at local, on-device agent workflows rather than cloud deployment.

What's new

• Muse Glimmer is described on its Hugging Face model card as "a 30-billion-parameter causal language model with a dedicated perception encoder, distilled from Muse Spark and purpose-built for autonomous agentic tasks on consumer hardware."
• The model targets end-to-end agentic task completion, reliable tool use, multi-step reasoning over long horizons, failure recovery when a tool call errors out, and multimodal input through its perception encoder, letting agents read screenshots, charts, and documents alongside text.
• It supports "controllable effort" (adjustable reasoning strength for a quality/speed tradeoff), works across agent scaffolds including OpenClaw and Hermes Agent, and is trained on data from more than 100 languages.
• Architecture specifics published with the release: a dense causal transformer with 29.6B total parameters, 52 layers, grouped-query attention (32 query heads / 2 KV heads), and a repeating local/local/local/global attention pattern with a 2048-token sliding window.
• Weights are released under an Apache 2.0 license on Hugging Face, with developer documentation, and Meta says optimized integrations for llama.cpp, MLX, and ExecuTorch are coming "in the coming days" for local deployment on Mac and PC hardware.
• The model was distilled from Meta's larger Muse Spark model using logit distillation, then further trained on longer-context agentic data and refined with supervised fine-tuning combined with on-policy distillation and reinforcement learning.

Context

Muse Glimmer follows Meta's Muse Spark and Muse Spark 1.1 releases earlier in 2026, positioning Glimmer as the compact, locally-runnable counterpart to those larger cloud-hosted models. It arrives amid a broader industry push toward small, capable on-device agent models — Liquid AI shipped its LFM2.5-2.6B on-device agentic model earlier this month — as labs compete to put agentic capability directly on phones, laptops, and other hardware without requiring a network connection or cloud inference costs.

Why it matters

Most agentic AI today still depends on round-tripping requests to cloud-hosted frontier models, which adds latency, cost, and a hard requirement for connectivity. A 30B model that fits on a single consumer GPU and still targets full agentic workflows — tool use, multi-step planning, multimodal understanding — pushes that capability bar further onto local hardware, which matters for cost-sensitive deployments, privacy-sensitive use cases, and offline scenarios. Releasing it open-weight under Apache 2.0 also keeps Meta's open-source strategy intact even as its frontier labs increasingly split effort between large cloud models and smaller distilled ones aimed at developers building local-first agent products.