en.Wedoany.com Reported - On August 10, U.S. tech company Meta released the Muse Glimmer artificial intelligence model and opened its model weights under the Apache 2.0 license. The model, featuring 30 billion parameters, is designed for continuously running local agentic workflows and can be deployed on Macs or personal computers equipped with a single consumer-grade GPU, without requiring a persistent cloud connection.

Muse Glimmer supports text and image inputs, with training data covering over 100 languages. It can perform multi-step tasks such as function calling, code writing and debugging, and file and screenshot processing. The model can identify issues after a tool call fails and retry, and can adjust its reasoning intensity based on task requirements, allowing a choice between processing speed and output quality.

To adapt to consumer-grade hardware, Meta employed approximately 4-bit quantization technology, compressing a model that originally required over 55GB of memory to under 20GB, enabling the model weights, working memory, image perception encoder, and speculative decoding module to run simultaneously in environments with 24GB or 32GB of memory. The accompanying DFlash speculative decoding module can generate candidate tokens in batches; Meta testing shows it improves decoding speed by approximately 3.1x on an RTX 5090 GPU, and by 1.8x and 1.5x on M5 Max and M4 Max devices, respectively.

The model was trained using Meta's previously released Muse Spark as the teacher model, acquiring agentic capabilities through output distillation, long-context training, supervised fine-tuning, and reinforcement learning. Muse Glimmer model weights are now available on Hugging Face, and Meta will also provide adapted versions for frameworks such as llama.cpp, MLX, and ExecuTorch.

Glimmer is currently available to developers in the form of model weights and development documentation, and can be used to build personal agent applications such as schedule management, file organization, content writing, and local programming. On the same day, Meta listed local operation and private data processing as directions for "personal superintelligence," while its more powerful Muse Spark model weights remain closed.