Post-training Kimi K3 with Harvey for long-horizon legal work

This post details the post-training of Kimi K3 with Harvey's Tenet model for long-horizon legal work, demonstrating a 19.7% all-pass rate on the Legal Agent Benchmark (LAB) compared to base Kimi K3's 10.8%. It highlights the transfer of learned behaviors to unseen agentic benchmarks (Mercor's Apex Agents and Crosby's Redline Bench) with no significant regression on legal knowledge benchmarks. The post also touches upon the infrastructure requirements for large-scale reinforcement learning, emphasizing numeric alignment and batch invariance of the Fireworks platform.