The journal
Perspective / September 18, 2026 / 3 min read

A smaller kind of intelligence.

What if the most useful model was the one you barely noticed?

Start with one useful thing.

Not every interaction needs a general-purpose language model. Identifying a language, suggesting a topic, or recognizing a simple shape can be a narrow task with a narrow answer. That is the question behind Embermote: how much can a small, focused model do?

Keep the work close.

Local inference can remove a network dependency from an interaction. Once the model files are on the device, the input does not need to travel to a hosted inference service. That is a design property we can test, rather than a promise about every possible application.

Small is a constraint, not a benchmark.

A small file is not proof of a useful model. The first six Embermote models are deliberately simple research baselines. Our text corpora are small and authored; our audio and shape examples are procedural. These models make the runtime tangible, but their internal evaluations cannot establish real-world quality.

Build in the open.

Every model ships with its training recipe, an artifact checksum, measured bytes, and a description of what it cannot do. The next step is to evaluate on independent, appropriately licensed data and real devices before calling any model production-ready.

Explore the research models