Meta Releases Muse Glimmer : 30B Open-Source Multimodal Model for Local AI Agents

Meta has released Muse Glimmer, a new 30-billion-parameter multimodal model designed specifically for local AI agent applications. The open-source model, available now under the Apache 2.0 license, marks a significant push by Meta into privacy-focused, on-device AI that can handle both text and visual inputs without cloud dependency.

Muse Glimmer combines a 2-billion-parameter vision encoder with a 28-billion-parameter text decoder, using a hybrid attention architecture that alternates between sliding window and full attention layers. This design reduces memory requirements while maintaining strong performance across benchmarks. The model also ships with an optional speculative decoding drafter that can accelerate generation speeds, particularly for structured outputs like code.

On agentic benchmarks, Muse Glimmer achieves a 75.5 score on MCP Atlas and 74.6 on DeepSearch QA, outperforming comparable models like Gemma4-31B and Qwen3.6-27B in several categories. It also posts strong results on coding tasks — 76.0 on SWE-Bench Verified — and multimodal understanding tests including ScreenSpot Pro and OmniDocBench. Safety evaluations show a 28.4% attack success rate on Siren AgentDojo, indicating room for improvement in prompt injection resistance.

The model’s local deployment focus addresses growing demand for AI systems that operate without sending data to external servers. For developers building coding assistants, document analyzers, or personal agents, Muse Glimmer offers an alternative to cloud-based APIs with potentially lower latency and zero per-token costs. Its Apache 2.0 license permits commercial use and modification, making it suitable for both hobbyist projects and enterprise applications.

Hardware requirements remain substantial despite the parameter count optimizations. Running the full 30B model locally typically requires at least 24GB of VRAM with 4-bit quantization, or 48GB for 8-bit inference. The optional speculative decoding drafter adds memory overhead but can double generation speed for supported workloads. Meta has released day-zero support across major inference frameworks including transformers, llama.cpp, vLLM, and Hugging Face Inference Endpoints.

Community reaction has focused on the model’s agentic capabilities and privacy positioning. Early testers report competent performance on multi-step reasoning tasks involving tool use, though some note that the vision component struggles with complex spatial reasoning compared to specialized vision models. The release follows Meta’s broader strategy of open-sourcing AI models while competitors maintain more restrictive approaches.

Muse Glimmer enters a competitive local AI landscape that includes recent releases from Mistral, Qwen, and Google. Its distinctive combination of multimodal input, agentic tuning, and permissive licensing could make it a preferred choice for developers building offline-capable AI applications. The model is available now through Hugging Face and Meta’s AI research pages.

Looking ahead, Meta plans to release additional Muse variants optimized for specific use cases. The company has indicated that a smaller 8B parameter version is in development for broader hardware compatibility, while enterprise-focused editions with enhanced security features are expected later this year. For now, Muse Glimmer represents one of the most capable open-source options for local multimodal AI agents.