Meta open-sources Muse Glimmer, an AI agent for consumer GPUs

Meta Superintelligence Labs has open-sourced the weights for Muse Glimmer, a new 30-billion-parameter AI agent released under the Apache 2.0 license and built to run locally on a regular consumer GPU, according to Meta's AI Research blog. The model's weights are already live on Hugging Face. Meta says developers can use Muse Glimmer for local coding, function calling, running autonomous agents, and as an LLM-as-a-judge for evaluating other models.

The core idea behind the release is cutting AI agents loose from the cloud. Cloud-hosted models require constant network access and centralized infrastructure, which is a poor fit for workloads where an agent needs access to genuinely private data — schedules, messages, files on disk. Muse Glimmer, per Meta, is built for exactly that niche: the agent runs directly on the user's device, and sensitive data never has to leave it.

“We're releasing an open weight model, Muse Glimmer, a 30B agentic model with open weights under Apache 2.0. Muse Glimmer can run on 24GB of VRAM without losing agentic reliability.”

— Alexandr Wang, Head of Meta Superintelligence Labs

Quantized, the model shrinks to under 20GB, putting it within reach of mainstream gaming GPUs like the Nvidia RTX 5090, as well as Apple's M4/M5 Max chips. By Meta's own benchmarks, Muse Glimmer performs on par with or above similarly sized models — Gemma4-31B and Qwen3.6-27B — across a range of agentic, coding, and multimodal tests. The release landed alongside a lengthy programmatic essay from Mark Zuckerberg on Meta's open-AI strategy, plus an announcement that open weights for the larger Muse Spark 1.2 model are coming "in the coming weeks."

• Muse Glimmer: 30B parameters, Apache 2.0 license, weights live on Hugging Face
• Runs on 24GB of VRAM without losing agentic reliability, per Meta Superintelligence Labs' head
• Use cases: local coding, function calling, autonomous agents, LLM-as-a-judge

The bet on local AI agents is a logical extension of something we covered earlier: how Zuckerberg promised every user a personal AI agent within five years — an open model that fits on an ordinary gaming PC looks like one of the first concrete steps toward that goal.

This piece is informational, not a recommendation to buy, sell, or hold any asset.