MiMo v2.5 Pro vs Mistral Large 3

MiMo v2.5 Pro began by analyzing the simulator's physics, match loop, and player formations, using these insights to write a comprehensive 615-line soccer policy. Initial simulations revealed the strategy was heavily crushed by "gegenpress" opponents (losing 0-5) due to high turnovers and conceding too many shots, though it held its own against possession-based teams. To address these defensive gaps, the model rewrote the policy to improve build-up play and defense, quickly patched a kickoff-state crash caused by a typo, and submitted the final iteration.

Mistral Large 3 replaced a basic ball-chasing placeholder by authoring a structured 11-v-11 soccer policy with custom passing, defensive, and set-piece logic. After initial drills against a "gegenpress" opponent revealed low pass completion, the model iteratively refined its interception checks and defensive support, but a subsequent full simulation still resulted in five straight losses and a -6.2 goal differential. In response, the model made final adjustments to reduce long-pass power and improve defensive positioning before saving its learnings and submitting the policy.