The AI open-source ecosystem has seen a major update. Meta officially released its new open-source model Muse Glimmer recently, marking the company's first time since the release of Llama 4 to make model weights available to the public again.

As a dense multimodal model with 30 billion parameters, Muse Glimmer is deeply optimized for local and continuous agent workflows. The model uses a permissive Apache 2.0 license, supporting text processing and visual capabilities, and demonstrates strong performance in various complex multi-step reasoning and tool calling scenarios.

In terms of hardware compatibility and operational efficiency, the model shows high practical value. Through approximately 4-bit quantization, its size can be significantly reduced to below 20GB, allowing it to run smoothly on consumer-grade hardware (such as PCs or Macs equipped with high-performance GPUs) with 24GB or 32GB of VRAM. Combined with the DFlash block-level speculative decoding technology, the model's throughput on hardware such as RTX 5090 has been significantly improved, achieving several times performance acceleration.

In publicly released benchmark tests, Muse Glimmer has shown excellent overall performance, especially leading in core benchmarks such as agent tasks and MCP Atlas. Meta founder Mark Zuckerberg stated that the company will remain committed to the open-source approach and revealed that the weights of the latest base model Muse Spark 1.2 will also be gradually made available in the coming weeks.