The tuned 12B can think on hard choice questions
A third setting for the tuned 12B, commit-think, reads a few unsure choice questions again after a short reasoning step, and is right more often on hard decisions.
A tuned Gemma 4 12B: right more often on hard, rule-heavy decisions
Our tuned 12B is right about 7 points more often than the untrained 12B in the same layout on held-out hard decisions, and about 3 points more on everyday wording. Every way to run it.
decisio 0.27.0: truncation fixed under the header layout, and opt-in flavors
0.27.0 fixes --over-limit truncate under the header layout, and adds flavors, reading short requests twice and think temperatures, all off unless you ask for them.
decisio 0.26.0: the Gemma models read your questions first, and answers change
0.26.0 lists a request's questions before its text on the Gemma models, which answered more accurately on the 31B; anything taught keeps the old layout, and --layout plain restores it.
decisio 0.24.0: the 12B on a 32 GB card, batched serving, and thinking that reasons
0.24.0 adds the Gemma 4 12B in FP8 (full 65,536 tokens on a 32 GB card), faster opt-in batched serving, and a fix so think-when-unsure reasons.
decisio 0.25.0: taught tasks survive restarts, and replicas share them
0.25.0 saves taught tasks to disk so a restart keeps them, and shares them between replicas through object storage, so a task taught once reaches every server.
decisio 0.15.0: the 31B on a 48 GB Mac
0.15.0 adds an opt-in 4-bit build of the Gemma 4 31B for Macs with 48 GB or more. It costs 1.5 points of MMLU-Pro and is slow on long documents.
decisio 0.13.0: long documents, read whole
0.13.0 can read documents of about 98,000 tokens on the Gemma 4 31B and 65,536 on the 12B, as an opt-in. It is off by default, and the first question is slow.
decisio 0.11.1: the 12B no longer slows down when new documents arrive back to back
Two server changes: the Gemma 4 12B now registers a document's boundary before the question, and every request records how long it waited for the engine.
decisio 0.12.0: certified answers
0.12.0 lets you certify which answers to act on, from your own labelled examples. It also refuses non-JSON POSTs and sets memory limits on a Mac.
Fifty questions about one document, five systems, one box
How long five systems take to answer 50 questions about one document, measured in one session on one RTX PRO 6000. Time, not accuracy.