Chon-1.0-12B, an independent Gemma 4 LoRA, leads JevBench open weights

Source · Model card: Chon-1.0 (12B / E4B / E2B)

publicado (OK do Fabricio 21/10 BRT).

JevBench release v1.6.32 (20 October 2026) added three systems from the same draw: Chon-1.0-12B, Chon-1.0-E4B and Vega-U (Qwen3.5-4B reader). Chon-1.0-12B went straight to #1 of the Jev-class open-weight rows, with Capability 85.3 (Intelligence 79.9, Calibration 90.8). Next are basal-omni-1.0-ultra, Quyet-1.0-Large and Jeff-1.0-Large. It costs $0.050 per 1,000 decisions, 1.55 times Jev’s reference, which is inside the 2× cap. Its measured p50 is 0.134 s.

What it is. Chon is a set of LoRA adapters on Gemma 4 (12B, E4B, E2B) from independent developer Ninh Nguyen (@xoai), released under Apache-2.0 with a server that speaks TypeSafe’s /v1/systemone format. It answers choice, yes/no and score questions with a probability per option. On unsure questions it can take a short extra pass that generates text, so the author’s p95 is 23.5 s against a 0.11 s median.

Caveats. The board marks the 12B as a paid publication and lists its base model as undisclosed (the card says Gemma 4 12B). The E4B scores 68.4 but falls outside the Jev-class latency cap (p50 22.2 s). Chon is not on Decision Index 0.4 or Artificial Analysis yet, so this is one board’s result. The author’s “#1 on JevBench” holds for the Jev-class open-weights view, not the hosted API table.

Catalog page: Chon-1.0.