Legora CEO 谈法律 AI 选型：没有最强模型，靠路由和持续基准测试
Legora 的 CEO 说选法律 AI 别纠结哪个模型最强，他们直接做了套路由系统，按任务自动挑模型，窄任务还自己微调小模型。
法律 AI 公司 Legora 的 Max Junestrand 在伦敦接受访谈，回应“哪个模型最适合法律工作”这一常见问题。他的回答是没有最佳模型，前沿迭代太快，这个月的领先者下个月未必领先。Legora 的做法是为每个任务搭建路由基础设施，通过持续基准测试把任务分配给当前表现最好的模型。对于重复量大、答案唯一狭窄任务，他们选择后训练一个小模型，而不是依赖通用大模型。
@MaxJunestrand sat down with @etnshow in London last week to talk about one of the most common questions we get: which model is best for legal work?
His answer stays the same. There is no best model.
The frontier moves fast, and whichever model leads this month won't necessarily lead next month. So we built the infrastructure to route every task to whichever model handles it best right now, backed by continuous benchmarking. For the narrow tasks that repeat at volume and have one right answer, we post-train a smaller model instead of relying on a general one.
Thanks to @lukeknight and @ronanchamberss for the conversation.