Tacet α: a model that stays quiet

In a musical score, tacet tells a player to stay silent for a passage. That’s the idea behind our first model. Most language models are built to talk. Tacet α is built to decide, and then stop.

You give it the state of your app and a typed question. It gives back a probability for each allowed answer. No sentences, no JSON to repair, no answer outside the options you defined.

What we did

We started from Gemma 4 26B A4B, a mixture-of-experts model that uses about 4 billion parameters per token. We trained a small LoRA adapter, roughly 10 million parameters, on 8,400 questions split evenly across yes-or-no, choice, and score questions. Eight full runs fit on one H200 in under 21 hours.

Accuracy went from 74.2% to 80.5%. More importantly, the Brier score, which measures how wrong the probabilities are, dropped from 0.486 to 0.287. Adding a small KL penalty toward the base model cut drift in its general text behavior by 14% without costing accuracy.

Where it falls short

Deciding when to call a tool is still hard: on When2Call, Tacet α scores below the untouched base. And we haven’t yet measured latency or cost on matched workloads. Skipping generation helps, but the model still re-reads the state for every question. Sharing that work across questions is next.

The weights are open. Try it, break it, and tell us what you find.