Nagi-BIG, Jev, Laya and SemIf play Snake. Same seeds. Same sensors. Shrinking walls.
Every move below comes from an actual model call. No action overrides.
Arrow keys or buttons. One press = one move.
Casual human mode uses the same opening board and shrinking rule. Food respawns with browser RNG; your score is not part of the model comparison.
| Model | Round 1 | Round 2 | Round 3 | Points |
|---|
Cells show survival moves / food. 3–2–1–0 points by survival, then food. A seeded draw was predefined for exact ties; no draw was needed. Jev wins this pilot 8–7 over Nagi-BIG. Three seeds are not statistical evidence of overall model superiority.
Identical explicit sensors for each direction: collision, reachable space (flood fill), distance to food and next-ring safety. Models select UP / RIGHT / DOWN / LEFT. They do not see pixels. Runs are independent trajectories, not simultaneous head-to-head combat.
What latency meansLocal models ran serially on one H100; Jev ran through its remote API. Latencies are measured, but the paths and hardware are not matched. Simulation steps wait for decisions: this pilot measures survival quality, not a real-time timeout contest. Replay speed is compressed.
Built by Nagi's developers. This is an entertainment/research pilot, not an independent or official benchmark.
