The Jev recipe on a Gemma spine, at open-source velocity. GEV-26B-Decide is AutoTrust’s decision model on Google’s Gemma 4 26B backbone (A4B: 4B of it active at inference): image or text in, one probability score per allowed action out, one forward pass. 903,866 downloads and 1,588 likes in under a week on Hugging Face, Apache 2.0.
The base-family bet. The decision-model wave’s first generation all tuned Qwen backbones; GEV ports the recipe to Gemma and the same team ships JEV-27B-VL on Qwen with vision, so the two base families can be compared on the same task set. For deployment the Gemma port matters commercially: Gemma’s license terms are the lightest of the major bases for self-hosting at fleet scale, and the A4B active count means per-token compute is a quarter of the parameter count, so serving cost tracks roughly 4B-class even though 25.8GB of weights stay resident in 4-bit.
What runs where. 14.8GB of q4 weights plus a KV cache of 240MB per 1,000 tokens (30 layers x 8 kv-heads x 256 dim): a 24GB card serves routing-scale contexts comfortably, and a 384GB unified cluster (the 512GB Ultra tier) holds the whole 262k context with the KV ledger closed.