Meta published Muse Glimmer on 2026-08-10, a 30 billion parameter model released under the Apache 2.0 license and aimed squarely at local agentic use. It pairs a text decoder with a dedicated perception encoder so the model takes interleaved text and image input in one pass, and Meta describes it as distilled from Muse Spark and purpose-built for autonomous, multi-step tasks that run without cloud infrastructure or a network connection. The accompanying Hugging Face release describes the split as a 2 billion parameter vision encoder and a 28 billion parameter text decoder.
The hardware story is the point of the release. At full precision a 30 billion parameter model needs more than 55 GB of memory; Meta reports that 4-bit quantization brings Muse Glimmer under 20 GB, which puts it inside the 24 GB to 32 GB envelope of MacBooks and consumer GPUs. A speculative decoding method Meta calls DFlash adds a further 1.5x to 3.1x generation speedup depending on hardware. The model was trained on data covering more than 100 languages, and Meta benchmarked it against Gemma4-31B and Qwen3.6-27B across agentic, coding, multimodal, safety and reasoning suites. On the agentic benchmarks reported at launch it scores 75.5 on MCP Atlas and 74.6 on DeepSearch QA.
Distribution was set up for immediate use rather than staged access. Weights sit on Hugging Face under the meta-models organization, with GGUF quantizations and ExecuTorch builds published alongside the base repository, and day-zero support in transformers, llama.cpp, vLLM and Hugging Face Inference Endpoints.
The significance is less about the benchmark table than about where an agent can now live. A permissively licensed, tool-using, multimodal model that fits on a laptop changes the calculus for workloads that cannot leave a device or a network boundary, and it puts a genuine Apache 2.0 option back in Meta’s column at a moment when the strongest open-weight agentic models have mostly come from Chinese labs under bespoke licenses.