# Meta Open Sources Muse Glimmer: Apache 2.0, 30B, Runs on a Laptop | Alher Tech > Meta released Muse Glimmer under Apache 2.0: a 30B multimodal model with no revenue or user cap that runs on a single consumer GPU. What it changes for a business, and where the enthusiasm should stop. - Canonical page: https://alhertech.com/en/blog/meta-open-sources-muse-glimmer/ - Site: Alher Tech (custom software, AI agents and SEO engineering, https://alhertech.com/) - Contact: https://alhertech.com/en/contact/ --- On 10 August 2026 Meta released Muse Glimmer, a 30-billion-parameter multimodal model, under the Apache 2.0 licence, and said it will publish the weights of Muse Spark 1.2 as well. Mark Zuckerberg wrapped it in a 6,500-word essay arguing that advanced AI should not end up controlled by a handful of companies. Strip away the manifesto and something concrete is left for anyone running a business: a capable agent model that fits on a laptop, with a licence that does not put a ceiling on how much you use it. ## What Meta Actually Released Muse Glimmer is a 30B dense multimodal model with a dedicated perception encoder, built for agents that run continuously on a local machine rather than for chat. Meta quantised it to roughly 4-bit and got it under 20GB, and reports testing it on a MacBook M4-Max, an M5-Max and an RTX-5090. In other words, one consumer GPU or a well-specified laptop, not a cluster. The weights are on Hugging Face and the context window is 131,072 tokens. Alongside it, Meta says the weights of Muse Spark 1.2 are coming: the far larger reasoning model, with a one-million-token window, that accepts text, images, audio, video and PDFs. Glimmer is distilled from it, which is why a 30B model behaves better on agent tasks than its size suggests. ## The Licence Is the News, Not the Model Apache 2.0 matters more here than any benchmark. It is recognised by the Open Source Initiative and it allows commercial use, modification, redistribution and sublicensing with no revenue cap and no user cap. That is a break from how Meta shipped Llama. The Llama community licence carried a threshold at 700 million monthly active users, and although almost nobody in Europe was ever going to hit it, its existence was enough for legal teams to treat the model as conditional rather than free. Apache 2.0 removes the conversation entirely. It is also a reversal. Meta had moved away from open weights in April, when the proprietary Muse Spark succeeded the Llama family. Glimmer is the first fully open release since. ## What It Changes for a Business Three things become possible that were not, or were expensive: - Data that never leaves. A model running on your own hardware means customer records, contracts and medical or legal documents are processed without leaving the building. Under the GDPR that removes an entire category of problem: no international transfer, no processor to audit, no clause to negotiate. - Cost that stops scaling with usage. An API bills per token forever. A local model has a hardware cost and then near-zero marginal cost, which changes the arithmetic for anything high-volume and repetitive, such as classifying documents, extracting fields or reviewing text. - Latency without a round trip. An agent that reads local files, calls local functions and answers without leaving the machine is a different product from one waiting on a network hop, and it keeps working when the connection does not. ## Where the Enthusiasm Should Stop A 30B distilled model is not a frontier model, and Meta does not claim it is. Its comparisons are against Gemma4-31B and Qwen3.6-27B, not against Claude Opus or GPT-5.6. For the hardest reasoning and the most complex refactors the gap is real, and the 131K context is well short of the million-token windows the frontier tiers now offer. There is also a caveat that comes with every open-weight release: safety alignment can be fine-tuned back out. If you self-host, whatever guardrails you need are your responsibility, not the vendor's. And the second half of the announcement has not happened yet. Muse Spark 1.2's weights are promised, not published. Until they are, the open part of this story is the 30B model. ## Zuckerberg's Argument, and Who It Is Aimed At The essay makes a geopolitical case rather than a technical one. Zuckerberg argues that the risk is not open models but concentrated control of closed ones, naming OpenAI and Anthropic by implication, and frames Chinese open models from DeepSeek and Moonshot as the competition to beat. His line is direct: restricting access to foreign open source models is not the answer, and the goal should be for American open source models to be the best in the world. It is worth reading it for what it is. Meta does not sell model access the way OpenAI and Anthropic do, so open weights cost it less and hurt its competitors more. That does not make the argument wrong, and it does not make Apache 2.0 any less real. It does mean the manifesto and the licence should be judged separately. ## Should You Use It If you are running high-volume, repetitive language work on sensitive data, this is the first release in a while that is worth an afternoon of testing on your own machine. If you need frontier reasoning, it is not that, and the honest comparison is against other open models rather than against the closed frontier. The useful move is not to pick a side. It is to measure: run the task you actually have against Glimmer locally and against whatever API you use today, and compare quality, latency and total cost including the hardware. That answer is specific to your workload, and it is the only one that matters.