Meta has introduced Muse Glimmer, a new artificial intelligence model from its Superintelligence Labs that is small enough to run locally on a single consumer graphics card. The 30-billion-parameter model is being released with open weights under an Apache 2.0 licence, the company announced today.
According to Meta, the model is designed for always-on agent workflows, meaning it can handle tasks such as managing schedules, drafting messages, organising files, and executing multi-step tool commands without needing a cloud connection. It supports text and image inputs, handles function calling across extended workflows, and can recover from failed tool calls rather than stalling.
To fit on consumer hardware, Meta says it used quantisation to compress the model to roughly 4-bit precision, bringing its memory footprint below 20 GB. This leaves enough room for the model to run alongside its working memory and perception encoder within the 24 GB or 32 GB limits of a typical laptop or desktop GPU. Meta says the compression introduces little to no performance loss on agentic tasks and that the model is fast enough for real-time interaction on devices such as a MacBook with an M4 or M5 chip, or an Nvidia RTX 5090.
Muse Glimmer was trained using a distillation process that transfers agentic reasoning from a much larger teacher model. It was evaluated across several benchmarks, including DeepSearch QA, MCP-Atlas, and SWE-Bench, and Meta says it performs strongly against comparable models in its size class, such as Gemma4-31B and Qwen3.6-27B. The model supports more than 100 languages.
The model weights are available now on Hugging Face. Meta said integrations with Ollama, LM Studio, Unsloth, llama.cpp, MLX, and ExecuTorch will follow in the coming days. The company says it is also working with hardware partners including AMD, Arm, Dell, Intel, and Nvidia to optimise performance across devices.
