一篇把判别式决策模型搬到卫星通信上的论文,单次前向 90 毫秒出决策,准确率还比生成式基线高出一大截,做边缘部署的可以看看。
#Laya
The practical catch is that Laya's confidence formula differs from Jev's, so existing Jev confidence thresholds cannot be carried over — Unsloth advises writing rules on probabilities instead, such as triggering when technical probability exceeds 0.8. Teams running decision routing on constrained hardware or air-gapped environments get a drop-in local option with documented limits: 64 questions per request, 255 options per choice, and label truncation past roughly 20 described options. Named alternatives to compare against are the Jev API itself, since the migration path is designed to be nearly transparent.
The setup turns local-versus-API latency into a visible, playable comparison — a concrete way to feel how response delay shapes agent behaviour in real time. Builders interested in agent frameworks can clone the GitHub repo and adapt the same head-to-head harness to benchmark their own local models against hosted APIs. It also doubles as a showcase of Typesafe's Jev agent in an unusual adversarial role.
开源社区真牛,把 Laya 移植到 Apple MLX 上,性能优化后每秒能做 60 次决策,在 M2 上还只占 1G 内存,玩贪吃蛇都丝滑。