$500 RL Fine-tune of 9B Open Model Beats Frontier Models on Catalog Review
核心: GRPO fine-tuned Qwen3.5-9B achieves 87.3% on catalog review vs frontier best 76.9% (5 models: GPT-5.5, GPT-5.6-sol, Gemini 3.1 Pro, Claude Opus 4.8, Fable 5). Cost: 19-172/1k.
技术: 2x RTX PRO 6000, $500 GPU time, 3.5 days, prime-rl framework. Crossed frontier after only 250 steps (~1 day). Digital twin with 177,767 episodes. Asymmetric scoring (missed violation costs 7x more than false alarm).
案例: Bridgewater (30% fewer errors), Harvey (legal agent beats GPT-5.5 & Opus 4.8), Intercom Fin Apex, LinkedIn (4% more accurate, 75x cheaper), AT&T (17% better, 12x faster), Phonely (99.2% accuracy).
对我有什么用: Hermes could fine-tune 7-9B open models for specific tool-calling workflows. Digital twin approach applicable to skill QA. “Prompt tax” finding (28-55% token inflation for <5% gain) directly relevant to Hermes context engineering.