Meta Open-Sources Muse Glimmer, a 30B Model Built for On-Device Agents
Meta Superintelligence Labs released Muse Glimmer on August 10, 2026, an open-weight, 30-billion-parameter model designed specifically for running AI agents locally on personal devices without cloud dependency. Meta describes it as “built for agents that don’t stop,” with reliable tool-calling, persistent state across restarts, and self-managed memory.
Details
- Size and footprint: 30B parameters, runs on a single consumer GPU or Mac, with a memory footprint under 20GB when quantized
- License: open-sourced under the permissive Apache 2.0 license, with weights available on Hugging Face
- Multimodal: supports text and image input, and was trained on more than 100 languages
- Built for agentic work: tuned for end-to-end agentic task completion, reliable tool use and function calling, multi-step reasoning across extended workflows, and failure recovery/error diagnosis
- Benchmarks: Meta says Muse Glimmer outperforms comparably sized models, including Gemma4-31B and Qwen3.6-27B, on agentic, coding, multimodal, safety, and reasoning benchmarks, with strong results on DeepSearch QA, MCP-Atlas, τ-Bench, and SWE-Bench
- Speed: speculative decoding delivers up to 3.1x faster generation on an RTX 5090, 1.8x on an M5 Max, and 1.5x on an M4 Max
- Training approach: distilled from Meta’s larger Muse Spark via logit distillation during pretraining, then put through long-context and agent-focused training, supervised fine-tuning, reinforcement learning, and further distillation
How to try it
- Download the weights now from Hugging Face
- Runs today with vLLM, Ollama, LM Studio, SGLang, and llama.cpp; native llama.cpp, MLX, and ExecuTorch integrations are rolling out within days
- Also available through Fireworks and OpenRouter
- Full announcement: research.meta.ai/blog/introducing-muse-glimmer-open-agentic-model