deck31b

deck31b serves frozen google/gemma-4-31B-it at revision 842da3794eaa0b77d5f08bae87a17459d91ff475. This repository does not contain those weights. The server downloads them and quantizes linear layers to FP8 e4m3 weights and activations at startup.

We did not train the checkpoint. The prompt and letter readout follow the public Cygnet recipe (blockbrain-ai/cygnet-recipe at 81974de878a7385f878be828205c02064f8bc045), with temperature 3.4.

Code: https://github.com/krishna-gogineni-765/deck31b

git clone https://github.com/krishna-gogineni-765/deck31b
cd deck31b
python -m venv .venv && source .venv/bin/activate
python -m pip install --upgrade pip
python -m pip install --index-url https://download.pytorch.org/whl/cu128 torch==2.11.0
python -m pip install -e .
deck31b-serve --port 8090

The server has no authentication and listens on 127.0.0.1 unless --host is set.

POST /v1/systemone supports noul, choice, and score. The base weights remain under the Gemma Terms of Use. This server code is Apache-2.0.