Muse Glimmer

A 30-billion-parameter open-weights agentic model under Apache 2.0 that runs locally on a single consumer GPU.

Handles local agents and function calling, local coding, and LLM-as-a-judge evaluation, with end-to-end task completion, multi-step reasoning, failure recovery, and multimodal input
Trained in three phases: pre-training by logit distillation from a larger teacher model, mid-training on longer-context agentic data, then supervised fine-tuning combined with reinforcement learning
30 billion parameters, with quantized versions compressed under 20 GB
Released under Apache 2.0, a permissive license that allows commercial use and modification
Runs on a Mac or PC with a single consumer GPU, validated on MacBook M4-Max, M5-Max, and RTX-5090, needing 24-32 GB of memory with quantization and speculative decoding
Available on Hugging Face at huggingface.co/meta-models/Muse-Glimmer-30B, with llama.cpp, MLX, and ExecuTorch integrations named as coming

Running agentic workflows and tool calling entirely on your own hardware, with no API calls leaving the machine

Data cannot leave your device, or you want agent capability without per-token cost

Teams with data-residency or confidentiality constraints that have blocked them from using hosted agents

Join the newsletter to get ahead with AI.

I share strategies, new modules, free guides, and tools to get you better at using AI in your work.