Meta releases Muse Glimmer, a 30B AI agent that runs on your GPU, not their servers

Meta has built a 30-billion-parameter AI agent that runs entirely on consumer hardware — and released the source code for free. The model, called Muse Glimmer, does not route your tasks through Meta’s servers. It executes locally on your own GPU, and at a quantized memory footprint of 18 to 20 gigabytes, it fits within the envelope of current high-end consumer graphics cards.

That matters because the previous generation of local models — the 7-billion- and 13-billion-parameter variants that filled the Ollama library over the last two years — were capable text generators but poor autonomous agents. They struggled with multi-step reasoning, tool calls that failed midway, and tasks that required remembering what happened three steps ago. Muse Glimmer is specifically designed for those workflows: coding, function calling, schedule management, file organization, and multi-step task sequences that recover automatically when a tool call breaks.

The underlying architecture is a distillation of Muse Spark, the commercial model Meta began selling API access to last month. Through logit distillation and reinforcement learning, the company compressed its closed flagship into a form that fits on a 24-gigabyte GPU — the RTX 5090, Apple’s M5 Max, or the M4 Max. Speculative decoding pushes throughput further: the RTX 5090 runs the model 3.1 times faster than without it, the M5 Max at 1.8 times, the M4 Max at 1.5 times. In practice, that difference is the line between an agent that feels responsive and one that requires patience.

The model processes interleaved text and images — screenshots, documents, mixed-content inputs — and operates across more than 100 languages. It connects natively to the frameworks that local AI users already run: Ollama, LM Studio, llama.cpp, MLX. If you already pull models from Hugging Face and run them on your hardware, Muse Glimmer drops into that workflow without additional tooling.

The license is Apache 2.0, which means there are no commercial restrictions. Individuals, companies, and researchers can use it without paying Meta. That puts it in a different category from OpenAI‘s and Anthropic‘s models, which require paying per token even for development work. Against Google’s Gemma4-31B and Alibaba’s Qwen3.6-27B — its closest open-weight competitors by scale — Meta positions Muse Glimmer as strong for its size on agentic and general language-model benchmarks, without making more specific claims about which tasks it wins.

The hardware threshold remains selective. Fitting 18 gigabytes of model weights onto a GPU sounds accessible until you calculate what GPU that requires. An RTX 5090 currently costs above two thousand dollars. Apple’s M5 Max machines start near three thousand. This is local AI for a specific income bracket, not a replacement for the cloud for most users. The quantized version is also a reduced-fidelity sibling of Muse Spark: it inherits the training signal from the commercial model, but it is not the commercial model. Meta continues to sell access to the higher-capability version.

What changes at this release is the capable agent tier of local models. Running a 30-billion-parameter model trained for autonomous task completion is different from running a chat model at the same parameter count. The distinction is in how the model handles tool failures and multi-step sequences — not in how well it completes a single sentence.

The weights are available on Hugging Face from August 10, under Apache 2.0. Meta has confirmed device-level optimization partnerships with AMD, Arm, Dell, Intel, and Nvidia, with further tuned builds for specific hardware configurations to follow. The local AI community will stress-test the model within days, as it does with every significant release — that community’s benchmark results will provide a cleaner picture of where Muse Glimmer actually stands within two weeks of today.

Tags: meta, AI agents, Local AI, open-source AI, GPU, Meta Superintelligence Labs