Meta AI Unveils Muse Glimmer: A 30B Open-Weight Agentic Model for Consumer GPUs

Meta has 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. At full precision, a 30B model needs over 55 GB of memory, but Meta compresses it to roughly 4-bit precision and adds block-level speculative decoding, allowing it to run on one consumer GPU or a Mac with no network call. The Hugging Face collection includes BF16 weights, GGUF k-quants, ExecuTorch builds, and the DFlash drafter. Muse Glimmer is a dense causal transformer with a perception encoder, using grouped-query attention with 32 query heads and 2 KV heads, and a 131,072+ context length. Training involved logit distillation on Muse Spark's outputs, mid-training with longer-context agent-heavy data, and post-training with supervised fine-tuning and reinforcement learning. The compressed version takes under 20 GB for the language model, fitting in 24 GB or 32 GB envelopes. Two quantized builds are available: K-Quant-Dynamic targets 32 GB VRAM with 0.2% average degradation, and K-Quant-17GB targets 24 GB VRAM with 1.0%. DFlash, a block-diffusion drafter, predicts 16 tokens in one forward pass, achieving a 3.1x speedup on RTX 5090. Benchmarks show Muse Glimmer leads on MCP Atlas (75.5), DeepSearch QA (74.6), Gaia2 (43.3), SWE-Bench Pro (51.2), and AIME 2026 (94.7), but trails Qwen3.6-27B on OSWorld-Verified (65.9 vs 75.6) and TerminalBench 2.1. On safety, Siren AgentDojo attack success rate is 28.4 with utility 94.2, and Meta rates chem/bio, cyber, and loss-of-control risk at moderate or lower.