Meta open-sources Muse Glimmer, a 30B AI model built to run offline on consumer PCs

Meta ↘ 0.86% has released Muse Glimmer, a 30-billion-parameter open-weight AI model that can run entirely offline on consumer hardware, marking a notable step in the company's push to make capable AI accessible without cloud subscriptions or internet connectivity.

A Local-First AI Model

Released under the Apache 2.0 license, Muse Glimmer is available on Hugging Face for anyone to download, modify, and deploy freely. The model accepts both text and image inputs, supports over 100 languages, and handles context windows exceeding 131,000 tokens.infoq+1

At full precision, the model would require roughly 64 GB of video memory — far beyond typical consumer hardware. Meta addresses this through 4-bit dynamic quantization, compressing the model into two variants: a 32 GB version that loses just 0.2% accuracy, and a 17 GB version that fits within 24 GB of VRAM with about a 1% accuracy trade-off. Compatible hardware includes Nvidia's RTX 5090, RTX 4090, RTX 3090, or Apple Silicon Max chips.tech.yahoo+1

A speculative decoding system called DFlash pairs a lightweight "drafter" model with the main network, proposing multiple tokens at once for parallel validation. On an RTX 5090, this pushes generation speed from 74.9 tokens per second to 233.4 — a roughly 3.1x throughput increase.infoq+1

Agentic Capabilities and Benchmarks

Muse Glimmer is built for autonomous, multi-step task execution. When an API call or terminal command fails, the model diagnoses the error and attempts alternative approaches rather than halting. A dedicated 1.8-billion-parameter perception encoder processes screenshots, diagrams, and documentation inline during workflows.infoq

On standardized benchmarks including SWE-Bench, DeepSearch QA, and MCP-Atlas, Muse Glimmer leads peer models such as Google's Gemma 4 31B and Alibaba's Qwen 3.6 27B in planning and multi-step tool reliability, though it trails in desktop operation tasks. The model is distilled from Meta's larger Muse Spark through a multi-stage process combining logit distillation, long-context mid-training, and reinforcement learning alignment.tech.yahoo+1

Zuckerberg's Open-AI Strategy

The release arrives alongside a broader strategic push by CEO Mark Zuckerberg, who published a lengthy manifesto this week framing the effort as "superintelligence for everyone". Axios reported that Meta's open-weight models are designed to "squeeze the frontier from below," delivering competitive performance at dramatically lower cost than proprietary alternatives.axios

The pivot follows Meta's disappointing Llama 4 launch and a subsequent overhaul that included a $14.3 billion investment in Scale AI. Still, a Meta employee told Axios the company "still feels behind the frontier labs in model performance and development".axios

Muse Glimmer is available now through Hugging Face, Ollama, LM Studio, and other local inference frameworks.tech.yahoo+1