When a Powerful AI Costs Nothing Per Use
Meta just released a powerful AI that anyone can run on their own hardware for free, and that changes how much it costs to build AI-powered products.
- 30B
- parameters, distilled to fit a single GPU
- Apache 2.0
- license — free for commercial use
- 24 GB
- GPU memory to run it on your own hardware
- 51.2
- coding benchmark score (SWE-Bench Pro)
Running a capable AI used to mean paying an API bill every time it ran — fractions of a cent per chunk of text, and those fractions add up fast at scale. Meta released Muse Glimmer on August 10 under an Apache 2.0 license: a 30-billion-parameter model built for real agentic tasks, free to use commercially, and designed to run on a single consumer graphics card.
Muse Glimmer is built for the kind of work where an AI takes a goal, breaks it into steps, calls tools, and keeps going until it is done. It handles text and images, supports coding, document analysis, and function-calling, and fits on a 24 GB graphics card after compression. Meta reports a SWE-Bench Pro score of 51.2 — a benchmark that measures how well the model fixes real software bugs — which puts it in range of paid frontier models.
Apache 2.0 is the important part. Many 'open' AI models carry restrictions: no commercial use above a certain scale, no modifications. Apache 2.0 has none of those. For teams where data must stay on your own servers — in finance, law, or healthcare — this is the first time a model this capable has been available without strings attached.
The practical step is to run a cost comparison. Muse Glimmer will not beat the biggest paid models on every task, but for internal tools, coding assistants, or workflows that fire hundreds of times a day, running your own model may now cost less than paying per token. It is available now on Hugging Face and works with ollama 0.32.7 and later.