Meta AI Releases Muse Glimmer: A 30B Open-Weights Agentic Model That Runs on One Consumer GPU

Details

Meta's Muse Glimmer is a dense causal transformer distilled from the closed Muse Spark model, engineered specifically for local agent deployment on consumer hardware. The architecture uses 32 query heads and 2 KV heads with grouped-query attention, a [Local, Local, Local, Global] attention pattern with a 2,048 sliding window, and RoPE with theta 500,000 applied to local layers only. The full model requires over 55GB at BF16 precision. Meta addresses this with two quantized builds: K-Quant-Dynamic targeting 32GB VRAM at 0.2% average degradation, and K-Quant-17GB targeting 24GB VRAM at 1.0% degradation. A dedicated 1.8B ViT-G/14 perception encoder handles up to 4,096 visual tokens per image. Context length exceeds 131,072 tokens, vocabulary is 202,048 tokens, and the knowledge cutoff is January 4, 2026. DFlash speculative decoding is the key inference optimization: it drafts likely next tokens in parallel with the main model, then verifies them, yielding 3.1x faster token generation compared to standard autoregressive decoding. The drafter ships alongside quantized weights on Hugging Face, along with ExecuTorch builds for edge deployment. Training ran in three phases: pre-training via logit distillation on Muse Spark outputs, mid-training on agentic and long-context data with richer reasoning traces, and post-training combining supervised fine-tuning with on-policy distillation and reinforcement learning across general, reasoning, coding, and agentic domains.