2026-09-18 · America/Los_Angeles · 社区动态 · #7

I tried Jev-style decisions with local Qwen: same accuracy, 239 ms vs 368 ms

I liked the idea behind Jev, but I don't want to send all my data to someone else's server. So I made choosekit, a small TypeScript package for getting choices and probabilities from a model running in llama.cpp. I tested it with Qwen3.8 27B Q4 XL against Jev, using SemIf's 144-task benchmark https://github.com/TheoLeeCJ/SemIf . Both achieved 96.53% accuracy. Median response time was 239 ms for local Qwen and 368 ms for Jev's hosted API. Full results and setup are in the README. GitHub: https://github.com/NotXf1le/choosekit

热度 50.1 / 100；排名与评分保留该期记录。