多家企业通过结合开源模型、专有任务数据和强化学习技术，成功微调参数规模为9B的开源模型，在各自的专业领域内超越了先进的闭源模型，同时大幅降低了推理成本[1]。
Bridgewater Associates的微调模型在其专业任务上将错误率相比最强前沿模型降低了约30%[1]。法律科技公司Harvey开发的强化学习模型在法律任务上超越了GPT-5.5和Claude Opus 4.8[1]。客服平台Intercom推出的Fin Apex模型每周能够解决近200万个客户问题，同样以更低的成本完成这一目标[1]。
Companies including Bridgewater Associates, Harvey, and Intercom have demonstrated that fine-tuning open-source models using proprietary task data and reinforcement learning can surpass advanced closed-source alternatives while substantially reducing inference costs.[1] Bridgewater Associates' fine-tuned model achieved approximately 30% lower error rates compared to the best frontier models available.[1] Harvey's reinforcement learning approach enabled its model to exceed the performance of GPT-5.5 and Claude Opus 4.8 on legal tasks.[1]
The approach has proven effective across different domains. Intercom's Fin Apex model, for instance, resolves nearly two million customer inquiries per week while maintaining lower operational costs than existing solutions.[1] These results highlight how organizations can leverage cost-effective optimization techniques—reportedly requiring only $500 in fine-tuning investment for a 9-billion-parameter model—to achieve specialized performance gains without relying exclusively on expensive proprietary models.[1]