Meta releases Muse Glimmer 30B, a local multimodal model for agentic use cases

Meta released Muse Glimmer, a 30-billion-parameter open-source multimodal model under the Apache 2.0 license, designed for local agentic use cases such as coding assistants, document analysis, and personal agents.

The model supports both text-only and multimodal inputs, including images and video, and is intended for privacy-aware deployments where data does not leave the device.

Muse Glimmer was distilled from Meta’s larger Muse model and includes a 2-billion-parameter vision encoder paired with a 28-billion-parameter text decoder, alongside an optional speculative decoding drafter for faster structured generation.

Day-zero support is available in transformers, vLLM, llama.cpp, and Hugging Face Inference Endpoints, enabling deployment across NVIDIA, AMD, and Intel GPUs without code changes.

Benchmark results show Muse Glimmer-30B leading or competitive with comparable open models across agentic coding, multimodal reasoning, and general capabilities benchmarks, including SWE-Bench Pro, OSWorld-Verified, and Charxiv Reasoning.

The model’s architecture uses hybrid attention with a repeating pattern of three sliding-window layers followed by a full-attention layer, gated grouped-query attention to reduce KV-cache memory, and Q-K normalization with extra query scaling for stable attention logits.

The vision encoder processes images and videos into shared embeddings, using a 2-frame-per-second sampling rate capped at 96 frames, with timestamped video placeholders interleaved in the text prompt.