Meta returns to open source with Muse Glimmer, an Apache 2.0 licensed 30B parameter AI model optimized for agents — available now

Meta has returned to open source with Muse Glimmer, a 30-billion-parameter AI model designed to run autonomous agents directly on consumer hardware. This marks a departure from the company's proprietary Muse Spark model, which was introduced in April. Muse Glimmer is licensed under the permissive Apache 2.0 open source license, a significant change from Llama's community license that had restrictions on commercial use and redistribution.

The model weights are available on Hugging Face, and support is being rolled out through various platforms, including Ollama, LM Studio, vLLM, SGLang, Together AI, Fireworks AI, and OpenRouter. Meta is working with hardware manufacturers AMD, Arm, Dell, Intel, and Nvidia to optimize performance across devices. The model was trained around the sequence of operations an autonomous agent performs, such as formulating a plan, calling tools, interpreting results, and recovering from errors.

It can operate as a fully capable agent while interpreting screenshots, charts, and documents, and invoking external tools. The model offers different reasoning settings and supports multiple languages. The hardware aspect is crucial, as the full precision 30B model requires over 55GB of memory, but Meta has developed several quantized versions that can run on consumer machines with 24GB or 32GB of VRAM.