MiniMax M2.7: First Self-Evolving Model — SWE-Pro 56.22%, GDPval ELO 1495 (Highest Open-Source)

MiniMax·high signal

MiniMax M2.7 claims to be 'our first model deeply participating in its own evolution,' autonomously optimizing its own performance scaffolds and achieving a 66.6% medal rate across 22 ML competitions. Benchmarks: SWE-Pro 56.22% (matching GPT-5.3-Codex), VIBE-Pro 55.6% (near Opus 4.6 parity), GDPval-AA ELO 1495 (highest among open-source models), 97% skill adherence rate on 40+ complex skills (2,000+ tokens each). Production incident recovery times reduced to under 3 minutes in real-world engineering scenarios.