Torchcast Decision 12B

One more voice in the flood, with the strictest license of the wave. Torchcast Decision 12B is an 11.96B-parameter decision model on the Gemma 4 Unified backbone from torchcast-ai, released October 2, 2026: the same typed-answer contract as its peers (a fixed set of options, one pick, a confidence number), 6.8GB in 4-bit so it lands comfortably on a 16GB card.

The license is the decision. CC BY-NC 4.0 means non-commercial: research, evaluation, hobby rigs, and coursework can run it; anything a business charges for cannot, not without a separate commercial license. Every other open row in the October decision wave (Clef, pplx-decider, GEV, JEV-VL) shipped Apache 2.0, so this is the card whose fine print a deployment checklist catches. The community quantizers published GGUF builds anyway (9,647 downloads on mradermacher’s 12B quant within days); nobody shipping it into a product can keep them there.

Where it sits. Compared with its Apache siblings at 2-4x the size, the interesting question is whether the 12B class earns its memory at all: the honest use is as an evaluation baseline (does a twice-as-small Apache model match it on your routing set?) or as a personal-standalone classifier. Not the default recommendation for production routing.