Meta launches Muse Glimmer, a 30B AI model designed for local AI agents

HIGHLIGHTS

Meta’s Muse Glimmer is a 30-billion-parameter open-weight AI model designed for local agent workflows.

The model can run on consumer hardware, including Macs and PCs with a single GPU, using quantisation and speculative decoding.

Muse Glimmer supports coding, tool use, multimodal inputs and long-horizon tasks, with weights released under the Apache 2.0 licence.

Meta has introduced Muse Glimmer, a 30-billion-parameter AI model designed to run AI agents locally on consumer hardware. Mark Zuckerberg’s company is releasing the model weights under the Apache 2.0 licence allowing developers to download, modify and deploy the model for their own applications. It is aimed at tasks such as coding, function calling, tool use and multi-step agent workflows. Meta says it can run on a Mac or PC with a single consumer GPU, reducing the need for cloud-based infrastructure or a constant internet connection.

Muse Glimmer is for agentic AI

Meta says Muse Glimmer has been trained to handle longer workflows, including tool calls, reasoning, coding and failure recovery. The model can identify failed tool calls and attempt to correct them instead of stopping the task. It also supports the multimodal inputs allowing agents to work with text and images such as screenshots, charts and documents. The model has been trained using data from more than 100 languages.

Also read: Amazon Great Freedom Sale 2026: Best smartphone deals for under Rs 50,000 

Meta used outputs from its larger Muse Spark model during pre-training through logit distillation. The company then added longer-context and agent-focused training, followed by supervised fine-tuning, reinforcement learning and further distillation across reasoning, coding and agentic tasks.

Quantisation brings the model below 20GB

To run a 30B model at full precision would require more than 55GB of memory, according to Meta. The company has therefore used quantisation to reduce the model to under 20GB at roughly 4-bit precision. This allows it to operate within systems offering 24GB or 32GB of memory while leaving room for its working memory and image-processing components.

Meta is also using speculative decoding to improve response speeds. A smaller companion model proposes groups of tokens, which Muse Glimmer then verifies. Meta says this increased decoding speed by up to 3.1 times on an Nvidia RTX 5090, 1.8 times on an M5 Max and 1.5 times on an M4 Max.

Muse Glimmer is available through Hugging Face. Support for platforms and frameworks including Ollama, LM Studio, llama.cpp, MLX and ExecuTorch is expected to follow.

Ashish Singh

Ashish Singh is the Chief Copy Editor at Digit. He's been wrangling tech jargon since 2020 (Times Internet, Jagran English '22). When not policing commas, he's likely fueling his gadget habit with coffee, strategising his next virtual race, or plotting a road trip to test the latest in-car tech. He speaks fluent Geek.

Connect On :