Meta Superintelligence Labs released Muse Glimmer on August 10, 2026 — a 30-billion-parameter open-weight agentic model licensed under Apache 2.0 and optimized to run entirely on consumer hardware (24–32 GB of VRAM). The model is distilled from the closed Muse Spark frontier model and includes multimodal (text/image), tool-calling, and long-horizon reasoning capabilities required for always-on local agents.
Using 4-bit quantization, the full model compresses from 55 GB (full precision) to under 20 GB, leaving room for working memory and speculative decoding. Meta bundles DFlash block-diffusion speculative decoding, which achieves 3.1× token generation speedup on an RTX 5090, 1.8× on an M5 Max, and 1.5× on an M4 Max — fast enough for fluid agentic interaction without network calls. Weights are available now on Hugging Face; native support for llama.cpp, ExecuTorch, and vLLM launching in coming days.
For architects: this moves agentic AI from cloud-only to local-first. The 30B model beats Gemma4-31B and Qwen3.6-27B on agentic benchmarks (MCP-Atlas 75.5 vs. 54.2 and 62.5; DeepSearch QA 74.6 vs. peers in the 50s). It underperforms on pure terminal/computer-use tasks but is designed for tool orchestration, coding, and reasoning. It signals the commoditization of sub-frontier agentic models and supports Zuckerberg's "personal superintelligence" thesis.