Go System One uses Gemma 4 12B for finite-choice decisions, ordered scores and yes/no probabilities. We started with a fixed Gemma baseline, built a llama.cpp prototype to define the expected results and speed, then replaced it with a native Go runtime.
The current NVIDIA path loads hand-tuned PTX kernels directly from Go through the NVIDIA Driver API. It needs no CGo, llama.cpp or CUDA toolkit at run time, and reduced the pinned RTX 3060 request from about 521 ms in the first native version to a 72.80 ms median. Automatic two-field batches reached 12.01 entries/s at 100 entries; benchmarks record the workloads, revisions and precision trade-offs.
84.85% public-subset accuracy (196/231) with JevBench v1.4.0: easy 48/48, original 71/72 and hard 77/111, with no failed requests. Our results and raw evidence use one fixed runtime revision. This is public accuracy, not the official sealed-set composite score or leaderboard rank.
Go 1.26.2 or the version in go.mod is required. Model files are downloaded separately under the Gemma licence.
make setup
make artifacts-download ACCEPT_GEMMA_LICENSE=1 HF_TOKEN="$HF_TOKEN"
make run BACKEND=nvidia LISTEN=127.0.0.1:8080Open http://127.0.0.1:8080/go-system-one for Noul, Choice and Score questions. Select Batch decisions for the boolean/enum context interface.