You Can (Maybe) Run Meta's Latest AI Model Locally on Your Computer

On Monday, Meta announced its latest AI model: Muse Glimmer. The institution isn't needfully advertizing this exemplary arsenic its biggest and best, however. While Meta has plentifulness to opportunity astir Muse Glimmer's capacity potential, location are 2 cardinal trading points for the exemplary successful the company's eyes: The truth that it's open-weight, and that it's designed to tally locally connected "consumer" Macs and PCs—though possibly not connected your computer.

What is Muse Glimmer?

According ot Meta, Muse Glimmer is simply a 30-billion-parameter AI exemplary optimized for "always-on section supplier workflows." What that intends is Glimmer is designed to tally autonomously straight connected your machine, alternatively than outsourcing that processing to the cloud. Local AI is overmuch much unafraid than cloud-based, arsenic your information stays wrong your system. When you pat into unreality servers, you're astatine the mercy of whichever institution owns that network, perchance putting your information successful jeopardy.

While location are galore usage cases for a exemplary for illustration Muse Glimmer, Meta is emphasizing its usage for agentic AI. AI agents tin execute tasks connected your behalf and are increasing successful popularity, arsenic users and developers tin inquire their agents to do immoderate number of functions. Running agents locally is besides rather popular, as evidenced by OpenClaw's infamous year, but the astir demanding models besides require powerful hardware. If you're moving pinch a laptop, you mightiness not person the power basal to tally AI agents.

You May Also Like

That's portion of Meta's transportation here. The institution says Muse Glimmer was trained to equilibrium capacity pinch constricted hardware. Meta says it puts Muse Glimmer done 3 sets of training to execute this: a "Pre-Training," which trained Muse Glimmer connected Muse Spark's outputs; "Mid-Training," which trained the exemplary connected much "agent-heavy" data; and "Post-Training," which fine-tuned the exemplary crossed a number of parameters. Meta says Muse Glimmer tin grip end-to-end agentic task completion, multi-step reasoning, betterment erstwhile a instrumentality telephone fails, and multi-model input and reasoning, among different functions. It is besides trained connected information from complete 100 languages. The institution says the exemplary performs "strongly for its size class" against galore modular LLM benchmarks.

Credit: Meta

While Muse Glimmer isn't arsenic demanding arsenic different models, it apt won't tally good connected the mean laptop. Meta says that a 30-billion parameter exemplary would typically require much than 55GB of memory, but done "quantization techniques," Muse Glimmer tin usage little than 20GB. When you return its "working memory" into consideration, Muse Glimmer will apt usage 24GB aliases 32GB of memory. If you person a "Pro" MacBook aliases PC, that mightiness suit you well. But if you're for illustration me, your laptop has 16GB aliases little of RAM—not rather what Muse Glimmer requires. Still, it's a measurement guardant for section AI use. (For reference, Meta says it tested Muse Glimmer connected MacBook M4 Max, M5 Max, and pinch an RTX-5090 GPU.)

Meta says Muse Glimmer is "open," which intends Meta released the model's training weights. As such, not only tin anyone usage the model, but they tin tune it arsenic well. If you person circumstantial needs for your AI model, you tin tinker pinch Muse Glimmer to amended service your usage.

How Muse Glimmer compares to different AI models

There are a lot of AI models connected the marketplace these days, but Meta only has nonstop comparisons pinch 2 different models: Google's Gemma4-31b and Alibaba's Qwen3.6-27B. According to Meta's benchmarks, Muse Glimmer thumps some models successful the following 12 tests: MCP Atlas, DeepSearch QA, 𝜏³-banking, WildClawBench, GAIA2, SWE-Bench Pro, Sci Code, Charxiv Reasoning, IFBench, AIME 2026, AA-LCR, and Beam 128K. Gemma4-31b still has the crown successful 4 of the benchmarks, while Qwen3.6-27B leads successful eight.

Meta didn't supply akin comparisons pinch models from manufacture leaders, for illustration OpenAI, DeepSeek, Z.ai, aliases Moonshot AI. However, Artificial Analysis has a ranking of each open-weight models connected the market. As of this article, Muse Glimmer (high) is really successful 18th place. Moonshot AI's Kimi K3 (max) is successful the lead, pinch Z.ai's GLM-5.2 (max) successful second, DeepSeek's V4 Flash (max) successful third, and Kimi K3 successful (low) successful fourth. For reference, Gemma4-31b is successful 32nd, while Qwen3.6-27B is really successful 17th. There are galore variables here, but this leaderboard gives america a speedy glimpse astatine really Meta's latest exemplary ranks amongst the competition.

How to effort Muse Glimmer

You tin effort Muse Glimmer correct now by downloading the weights from Hugging Face. Meta says the exemplary will beryllium disposable successful the coming days from apps for illustration Ollama, LM Studio, and Unsloth.