This tune was more successful than anticipated... First 9B I've tested that was able to make an atari breakout game without bugs (zero-shot, q8)

                        arc     arc/e    boolq
Qwen3.5-9B-Fable-5-v1  0.624    0.806    0.891
           Qwen3.5-9B  0.553    0.712    0.892

As always big thanks to @nightmedia for the benchmarks :)

The data for this model was easily extracted, formatted, and masked for training with Teich

This qwen3_5 model was trained 2x faster with Unsloth and Huggingface's TRL library.