Google released Gemma 4 on April 2, 2026 under Apache 2.0 — their first fully permissive open model. The 31B dense model scores 89.2% on AIME and 80.0% on LiveCodeBench. The 26B MoE variant delivers 97% of that performance with only ~4B active parameters. Here's the full breakdown.
Moonshot AI's Kimi K2.5 deploys up to 100 parallel sub-agents, scores 76.8% on SWE-Bench Verified, and costs $0.60/$3.00 per million tokens. We break down the benchmarks, the PARL training technique, and what actually works.
