Jev 与文本生成模型的取舍
Jev 不生成自由文本。Diogo 用一段对比视频说明并行决策模型与逐字生成模型在任务形态上的差别。

Taro L. Saito 指出,像 Jev 一样,本地 LLM 引擎通过放弃文本/JSON 生成、专注于判断,在 DGX Spark 上实现了 40ms 推理,速度极快。
Just like Jev, to think that a Local LLM engine capable of inference in 40ms is realized simply by abandoning text/json generation and specializing solely in judgment. It's running on DGX Spark, but anyway, it's incredibly fast.