Meta Superintelligence Labs推出了Muse Glimmer，一款拥有30亿参数的开源AI模型，以Apache 2.0许可证开放权重[1]。该模型专为本地代理工作流优化，支持本地代理、函数调用、代码生成和评估等功能[1]。Meta领导人表示相信所有人都应该获得超级智能的访问权限[2]。
该模型可在单个消费级GPU的Mac或PC上运行，量化后显存需求约为20GB（4-bit精度），支持24GB或32GB显存的设备[1]。Muse Glimmer支持100多种语言[1]，已在Hugging Face发布[1]，并可集成到llama.cpp、MLX、ExecuTorch、vLLM、SGLang等多个框架中[1]。模型的性能已在MacBook M4-Max、M5-Max和RTX-5090等设备上进行了测试[1]。
Meta has released Muse Glimmer, a 30 billion parameter open-source AI model designed for local agentic workflows [1]. The model is distributed under the Apache 2.0 license and can run on consumer-grade hardware, including a single GPU on a Mac or PC [1]. The weights are available on Hugging Face [1].
The model supports over 100 languages and is optimized for local agents, function calling, code generation, and evaluation tasks [1]. When quantized to 4-bit precision, Muse Glimmer requires approximately 20GB of memory and can operate on systems with 24GB or 32GB of VRAM [1]. Meta's engineering team has tested the model's performance on various platforms including MacBook M4-Max, M5-Max, and RTX-5090 hardware [1]. The model integrates with popular frameworks such as llama.cpp, MLX, ExecuTorch, vLLM, and SGLang [1].
Meta's leadership has expressed the belief that everyone should have access to superintelligence [2], positioning this open-weight release as part of that vision. However, this approach has sparked discussion around the balance between providing broad access to advanced AI capabilities and managing associated security considerations [2].