Hopper 12B (trained)
A rank-32 LoRA on a frozen Gemma 4 12B IT, served unmerged with its own readout map. A state, a question and a fixed set of options go in; one forward pass returns a probability per option and generates no text. Apache-2.0, unlike the research-only 4B Hopper.
The weights live on the hopper-12b branch of HopitAI/hopper (adapter at revision 4b605f25), not on main, which holds the research-only 4B adapter; this page therefore links the branch instead of listing files. The maker says to keep the adapter unmerged: a bf16 merge changed 2.1% of top choices on 1,932 decisions. The readout applies position priors for 2- and 4-option menus plus one temperature (1.65) fitted on HopitAI's own menus; menus longer than 26 options go through a tournament shortlist over several passes. The same serving code also runs the frozen base with a readout map (no adapter). HopitAI's own scoring of the third-party Decision Index suite gives 47.52, submitted to that board and not yet merged; a JevBench measurement request is open. Trained without the RACE passages, so the adapter is Apache-2.0; Gemma's prohibited-use policy applies to the base, which is a separate download.
What it decides
- choice — picks one option from a set
- score — places the input on an ordered scale
- noul — answers a yes/no question with one probability
At a glance
| Parameters | 12B |
| Base model | google/gemma-4-12b-it |
| Maker | HopitAI |
| Released | 2026-10-02 |
| License | apache-2.0 |
Read more
This page was opened by System One for HopitAI, who can claim the organisation and take it over at any time.