Meta has launched a new open-source AI model that can run directly on a computer without relying on cloud infrastructure or a constant internet connection.

The company said its new Muse Glimmer model has been designed to run on Macs and PCs equipped with a single consumer GPU, enabling developers to build AI agents, coding assistants, and automation tools that operate entirely on-device.

Unlike many large language models that depend on cloud-based computing, Muse Glimmer has been optimised for local deployment, allowing AI applications to function offline while keeping data on users' own devices.

Meta said the 30-billion-parameter model has been released under an Apache 2.0 open-source licence, with model weights available on AI data repository Hugging Face alongside developer documentation.

The technology has been developed for “always-on” AI agents capable of handling tasks such as scheduling, drafting messages, organising files, and interacting with software through tool calling.

According to Meta, the model supports long-context memory, multimodal inputs combining text and images, multilingual capabilities across more than 100 languages, and can recover from failed tool calls by retrying actions rather than stopping altogether.

To make the model practical on consumer hardware, Meta said it compressed its 30-billion-parameter architecture using quantisation techniques, reducing its memory footprint from more than 55GB to under 20GB. The company also introduced speculative decoding technology to accelerate text generation without affecting output quality.

Meta said Muse Glimmer has been trained using a multi-stage process involving pre-training, longer-context mid-training and post-training that combined supervised fine-tuning, reinforcement learning and model distillation from its larger Muse Spark model.

The company added the model was evaluated under its Advanced AI Scaling Framework before being released and performs strongly against rival models of a similar size across a range of benchmarks covering coding, reasoning, and autonomous agent tasks.

Initially, Muse Glimmer will supprt local deployments through downloads from Hugging Face, with integrations for developer tools including llama.cpp, MLX, and ExecuTorch due in the coming days. Support for platforms such as Ollama, LM Studio, Together AI, Fireworks AI, and OpenRouter is also planned.

Meta said it is also working with hardware partners including AMD, Arm, Dell, Intel, and NVIDIA to optimise performance across different devices.

---