KartiOS/Karti-Small-Support-9B
Text Generation • Updated • 22 • 1
Small open models trained against verifiable rewards, and the datasets behind them.
Note RL on a code-computed reward: 0.706 vs 0.531 base on held-out support triage.
Note 106 multi-turn support episodes + 63-policy pack; reward computable in code.
Note Vision SFT: invented-identifier rate 0.378→0.023, ScreenSpot-v2 0.718→0.907.
Note NVFP4 build for Blackwell GPUs.
Note Tool-calling agent model: recipe and model card; weights pending.