Meta has just given AI developers a reason to shut off the cloud.

Muse Glimmer, the company's newest model, is small enough to run entirely on a single graphics card, no data center or server farm required. Unveiled this week, it's designed to live locally on a gaming PC or a well-equipped Mac, marking a notable shift in how Meta wants people to actually use AI.

Big Model, Small Footprint

Muse Glimmer packs 30 billion parameters, distilled down from Muse Spark 1.2, the closed flagship model Meta released just days earlier. At full precision, a model that size would typically need more than 55 gigabytes of memory. That's far more than any consumer GPU offers. Meta's engineers compressed the weights to roughly 4-bit precision, bringing the total footprint down to well under 20 gigabytes, small enough to comfortably fit on a 24GB or 32GB card, the kind many developers and gamers already own.

To keep things fast despite the heavy compression, Meta paired the model with a lightweight companion that drafts likely next steps for Muse Glimmer to quickly confirm, a technique known as speculative decoding. Hardware partners report the payoff is substantial: on a single high-end GPU, the model can generate around 20,000 tokens per second, more than three times faster than it would run without the shortcut.

An Agent, Not Just a Chatbot

Rather than optimizing purely for conversation, Meta built Muse Glimmer around "always-on" agent work: managing a calendar, organizing files, writing and debugging code, or working through multi-step tasks with minimal hand-holding. It handles a context window of more than 120,000 tokens and reads text alongside images like screenshots and documents. It can also maintain memory across hours-long sessions, picking up where it left off even after a restart. When a tool call fails, the model is designed to figure out what went wrong and attempt a different approach rather than getting stuck. It also covers more than 100 languages and lets developers adjust how much reasoning effort it applies to a given task. Because everything runs on the local machine, it keeps working without an internet connection.

Free to Download and Modify

Meta is giving the model away. Muse Glimmer's weights are available now on Hugging Face under an Apache 2.0 license, which lets anyone download and modify the model, even for commercial products, without paying royalties. Support for local-inference tools is on the way too: llama.cpp integration is expected within days, along with MLX and ExecuTorch, and platforms including Ollama, LM Studio, Together AI, Fireworks AI, and OpenRouter already support it at launch. For larger-scale deployments, Meta points developers to vLLM and SGLang.

Chipmakers are already on board too. NVIDIA says the model runs across its GeForce and DGX hardware, plus Jetson. Meta is also optimizing performance with AMD, Arm, Dell, and Intel.

Small but Competitive

On benchmark tests, Muse Glimmer holds its own against other compact open models, edging out rivals like Gemma4-31B and Qwen3.6-27B on agent-focused evaluations and posting a strong score on a recent advanced math reasoning benchmark. It doesn't win across the board. Competing models still lead on some coding and computer-use tests. But for something that runs on a single graphics card, the results land surprisingly close to systems many times its size. Meta also says the model doesn't cross its internal bar for what the company defines as "frontier" AI, rating the chances of serious chemical, biological, cyber, or loss-of-control harm as moderate rather than high.

A Bigger Bet on Openness

The launch arrived alongside a lengthy essay from CEO Mark Zuckerberg making the case for distributing powerful AI widely rather than concentrating it inside a handful of companies' data centers, arguing that putting models directly into people's hands gives individuals more control over the technology shaping their lives. It's a notable pivot: Meta had pulled back from open-sourcing its most capable models for over a year, a stretch in which rivals such as OpenAI and Anthropic were widely seen as pulling ahead. Muse Glimmer, along with reported plans to eventually release weights for the larger Muse Spark 1.2 too, signals a return to that strategy, and echoes moves by Chinese labs like DeepSeek, which have built reputations on capable, openly licensed models that run on local hardware.

For a company that still earns most of its money from advertising, giving the model away may be less charity than strategy. Treating the model itself as a commodity could squeeze competitors whose businesses depend on metered cloud access. Investors seemed receptive to the idea, with Meta shares climbing nearly 3% in premarket trading Monday.