Jev beats Ollaya on DGX Spark in decision-AI race

The test was built as a game-like simulation titled "Kakeagare Irohazaka," set around Seiseki-Sakuragaoka Station in Tokyo's Tama area, where a steep climb called Irohazaka offers a natural route-choice challenge. Two decision AIs started from the station and, at each fork, received only minimal information — road conditions and slope for the next three or four branches and the straight-line distance to the goal — and had to choose a direction on the spot rather than plan a full route upfront. Basic conditions such as "distance priority" or "avoid slopes" could be set before the start.

The results varied with model weight, but Jev reached the goal first in every case. Ollaya's lightest winnow variant, 4.6B, wandered more and, under "avoid slopes," took nearly twice Jev's time and distance. The 7.5B version often made large detours, and the 12B version improved but sometimes climbed the slope outright under the slope-avoidance setting. Jev's average API processing time stayed stable at around 0.3 seconds, while Ollaya ranged from about 0.7 seconds to nearly 3 seconds depending on model size — and because Jev's timing includes internet communication, its inference speed appears markedly high.

Still, Ollaya with the 12B winnow model was judged superior on total distance, API calls and estimated travel time, making fewer but more careful decisions. The practical takeaway is that the open option can serve route-planning uses where a wait of two seconds or more is not a problem, and it costs nothing to run continuously, whereas Jev is the safer bet when large volumes of judgments must be made quickly and accurately.