Seeking Pareto optimality, whatever the scale
Visual RL earns 67.7% of the judge’s votes against the untouched Qwen model.
Jev scores 0.763 at 1536 questions and 0.763 at 2048, versus 0.815 for Voyage 4 Large on SICK-R.
Jev makes 24/24 objects pass their schemas without generating text.
Nimble 9B scores 95% to Jev’s 90%. The fastest local model, Tev1 0.8B, takes 219 ms to Jev’s 210 ms.
Encrypted compaction uses 22% fewer tokens than text summaries.
Jev scores 1.43× GPT-6 Luna’s F2 on PII masking.
Thinking solves 3.3× as many AIME problems as Jev step selection.
Exo made zero self-edits and passed zero tasks.