A small model built for workflows can be checked, measured and trusted.
In 2026 we stopped asking how big AI could get, and started asking how small it could be and still do the job.
More from the blog
How we get our models running at full speed on Intel Xeon
Our own quantization and the Argon engine take an open 27B model to 30 tokens a second, on CPU.
Why general-purpose AI wastes most of what you pay for
Most automation needs a small part of what a general model can do. You pay for all of it.