On August 10th, Meta released Muse Glimmer, a 30-billion-parameter open model built specifically to run local AI agents on consumer hardware — scheduling, file organization, tool calling, multi-step tasks — without shipping any of it to Meta's servers. Weights are on Hugging Face under an Apache 2.0 license, so there's no commercial restriction on using it.
Why this one is different from the last wave of "local" models
The 7B and 13B models that filled up Ollama's library over the last couple of years were solid text generators, but weak agents. Multi-step reasoning fell apart, tool calls failed halfway through, and they'd lose track of what happened three steps back in a task.
Glimmer is Meta's answer to that gap specifically — it's a simplified, efficiency-focused derivative of Meta's larger Muse Spark 1.2 model, purpose-built for "always-on" agentic workflows: the kind of thing that needs to keep running continuously on a personal machine rather than answering one prompt at a time.
What it actually needs to run
30B parameters, quantized memory footprint of 18–20 GB — within reach of current high-end consumer GPUs, no server rack required
Handles interleaved text and images (screenshots, documents, mixed content) across 100+ languages
Native support for the frameworks local-AI users already have installed: Ollama, LM Studio, llama.cpp, MLX, vLLM, SGLang — if you already pull models from Hugging Face, this drops into your existing workflow with no extra tooling
Meta is also working with AMD, Arm, Dell, Intel, and NVIDIA on hardware-specific performance tuning, and providing documentation for building custom agent scaffolds on top of it
The bigger picture
Glimmer didn't launch alone — Meta paired it with a 14-page essay from Mark Zuckerberg ("The Future Is for Everyone") arguing against AI capability staying locked inside a handful of companies, and confirmed weights for the larger Muse Spark model are coming too. It's a direct response to the pressure open models are putting on the market: Chinese labs — Moonshot's Kimi K3, Alibaba's Qwen3.8-Max, DeepSeek's V4-Flash — are shipping performance that rivals closed US frontier labs, and open weights are consistently cheaper to run at scale, which matters to anyone watching their own inference bill.
Where this fits in a self-hosted setup
For anyone already running local models as part of a homelab — Ollama containers, a GPU passed through to a VM, that kind of setup — Glimmer is one of the first local models actually aimed at running unattended agent workflows instead of just answering chat prompts. That's a meaningfully different use case from "local chatbot," and worth testing against real tasks rather than benchmarks.
I run local AI as part of my own homelab, so I'll be testing Glimmer against real agentic tasks rather than benchmarks — full writeup on El Rack once I've put it through its paces:
👉 Muse Glimmer, el nuevo modelo de Meta
(Spanish-language site — translation tools handle it cleanly if you don't read Spanish.)
I write about homelab, self-hosting, and local AI at El Rack — real testing from inside my own homelab, not just a spec sheet rewrite.
Top comments (0)