Meta releases Muse Glimmer: local, agentic, multimodal, and open-source AI model
Meta
Meta has released Muse Glimmer, a 30B parameter open-source multimodal AI model designed for agentic tasks, with day-0 support in Hugging Face transformers, llama.cpp, and vLLM. The model, featuring a hybrid attention architecture and a 2B vision encoder, is available on the Hugging Face Hub and demonstrates strong performance on agentic and coding benchmarks.
Meta has introduced Muse Glimmer, an open-source 30B parameter multimodal model, with day-0 support in transformers, llama.cpp, vLLM, and other libraries. The architecture combines a 2B ViT-style perception encoder for vision with a 28B text decoder, using hybrid attention with sliding window and full attention layers, gated grouped-query attention to reduce KV-cache memory, and Q-K normalization. It also includes an optional speculative decoding drafter for faster generation, particularly suited for coding. Benchmark results show strong agentic performance on tasks like MCP Atlas, DeepSearch QA, and SWE-Bench, as well as multimodal capabilities. The model is available on the Hugging Face Hub, and integration examples are provided for text-only inference, multimodal tool calling, object detection, and video inference using transformers and llama.cpp.
- Abbreviations
- VLM = Vision-Language Model — визуально-языковая модель
- KV = Key-Value — ключ-значение
- RoPE = Rotary Position Embedding — вращательное позиционное кодирование
- NoPE = No Positional Embedding — без позиционного кодирования
- RMS = Root Mean Square — среднеквадратичный
- GELU = Gaussian Error Linear Unit — гауссовская ошибка линейного блока
- MLP = Multi-Layer Perceptron — многослойный перцептрон
- CUDA = Compute Unified Device Architecture — архитектура унифицированных вычислительных устройств
- ROCm = Radeon Open Compute — открытые вычисления Radeon
- XPU = eXtreme Processing Unit — экстремальное процессорное устройство
Source: Hugging Face blog —
original