Muse Glimmer 30B

Muse Glimmer is a ~29.6B dense causal transformer with a ~1.8B-parameter ViT-G/14 perception encoder, released by Meta Superintelligence Lab in August 2026 under Apache 2.0 and distilled from Muse Spark. It is built for autonomous agent work that runs entirely on a local machine: tool calling, multi-step reasoning, failure recovery, and multimodal input (interleaved text and images, so an agent can read screenshots, charts and documents). 52 layers, hidden dimension 6656, GQA 32/2, a repeating [local, local, local, global] attention pattern with a 2048-token sliding window, and a context length of 131,072+. Meta ships two quantisations aimed squarely at consumer cards: K-Quant-17GB targets a 24 GB envelope and K-Quant-Dynamic targets 32 GB, with the full BF16 weights needing 64 GB. The release also includes a DFlash block-diffusion drafter that proposes 16 tokens per forward pass for speculative decoding; Meta measured 74.9 tok/s without it and 233.4 tok/s with it on an RTX 5090 under llama.cpp. Supported by mainline llama.cpp (arch 'muse-glimmer', with mmproj files for vision) and available from the Ollama library as muse-glimmer:30b.