Meta has released the weights for Muse Glimmer, a 30-billion-parameter model from its Superintelligence Labs that runs entirely on a laptop or a single consumer GPU. It is the first MSL model released with open weights, under a permissive Apache 2.0 licence — anyone can download it, modify it and run it.
Muse Glimmer joins a line of open releases Meta has shipped over the years, but this one is aimed squarely at agents rather than chat.
What Muse Glimmer can do
The model is built to work through problems, not just answer them. A developer hands it a job and a list of tools; it plans, executes, checks its own results, and recovers when a step fails. It accepts images and video as input, and was trained on data from more than 100 languages.
Runs on your own hardware
Muse Glimmer is distilled from Meta’s larger Muse Spark: a 28-billion-parameter text decoder plus a 2-billion-parameter perception encoder. Quantised to roughly 4-bit precision, it fits in under 20GB, leaving room in a 24–32GB memory envelope for working memory, image understanding and a speculative decoding drafter based on DFlash. Meta says the drafter speeds up generation by 3.1x on an RTX 5090, 1.8x on an M5 Max and 1.5x on an M4 Max.
Target hardware includes a Mac Mini or MacBook with an M4 or M5 chip and 32GB or more, an NVIDIA RTX 5090, or an AMD Radeon AI PRO. Launch support spans llama.cpp, MLX, Ollama, LM Studio, vLLM, SGLang, Together AI, Fireworks AI and OpenRouter, with NVIDIA, AMD, Intel, Arm and Dell optimising across devices.
On agentic benchmarks, Meta puts Muse Glimmer ahead of similarly sized open models: 75.5 on MCP Atlas against 54.2 for Gemma4-31B and 62.5 for Qwen3.6-27B, and 51.2 on SWE-Bench Pro. Meta also said it will open the weights for a version of Muse Spark 1.2 in the coming weeks.
Why Meta is going open
The release arrives with an essay from Mark Zuckerberg arguing that superintelligence should be distributed rather than centralised.
“Rather than centralizing superintelligence, we should distribute it widely and give every person the ability to direct it.”
Engadget notes that Meta’s Muse Spark is widely seen as weaker than rivals from OpenAI and Anthropic; open-sourcing the smaller model looks like a bid for users who prefer to run AI locally, on their own hardware and their own data. Meta’s first coding agent, Muse Code, runs on Muse Spark 1.2 — the same family now arriving with open weights, months after Muse Spark launched across Meta’s apps.
Muse Glimmer is free to download now from Hugging Face, with documentation at dev.meta.ai. It needs no cloud connection and no subscription — just a machine with enough memory.
What is Muse Glimmer?
Muse Glimmer is Meta’s first open-weights model from Superintelligence Labs: a 30-billion-parameter dense model designed for local agent workflows. It is released under a permissive Apache 2.0 licence, so anyone can download, modify and run it, including for commercial use.
What hardware do I need to run Muse Glimmer?
Meta targets a Mac Mini or MacBook with an M4 or M5 chip and 32GB or more, an NVIDIA RTX 5090, or an AMD Radeon AI PRO. The 4-bit quantised model fits in under 20GB inside a 24–32GB memory envelope.
Is Muse Glimmer free?
Yes. The weights are free to download from Hugging Face under an Apache 2.0 licence, which also permits modification and commercial use. The model runs locally with no cloud connection and no subscription.
Leave a Reply