The license is the first thing I checked. Apache 2.0, no asterisk below the line. Meta had not done that in a long while, and some open models have shipped with licenses where Europe simply was not on the list.
Fact
The facts: on August 10 Meta Superintelligence Labs released Muse Glimmer - a multimodal model with 30 billion parameters (a 29.6 billion text decoder plus 1.8 billion for perception), under an Apache 2.0 license. Context is 131,072 tokens, languages over 100. Per Meta data the model is built for agent work: tool calling and failure recovery in one and the same model. Weights are on Hugging Face, the package on Ollama, with quantized versions for home hardware and a variant for mobile devices.
Built for agents means a different animal from the chatbot. A chat answers you. An agent does work: calls tools, reads the result, recovers when something blows up. Until now those habits were glued on top with code. Here they are trained into the model itself.
For a small European company the clean license is not a legal detail but the difference between a toy and a tool. A model you cannot legally use is a model you do not have.
Openness at this scale is not a gesture. It is strategy: weights buy back an ecosystem.
We will try it on our own hardware and then I will tell you whether it holds. Promises in a model card do not count.
Official primary sources
Hugging Face Blog - Muse GlimmerOllama Blog - Muse GlimmerMeta Developers - Muse GlimmerThe visual is generated code-art, no third-party images.