WHAT NEW IN X11.01 SUN

AI collections from X · 22 collections · 64,204 posts

jevIMAGE

Benchmarking Jev and open-source Jev-like models on RAG agent routing

The author built a RAG agent routing benchmark of 100 multi-turn questions: Gemma 4 31B scored 77.0%, Jev 61.4%, djev-spark 32.2%, SemIf 24.0% and Laya 322M 0%; Jev reached 90.6% on deciding where to find answers but was weak at multi-turn state updates, with a p50 of 749ms.

大數軟體LargitData
@LargitData1Archived 2026.09.21Originally posted 2026.09.20
Benchmarking Jev and open-source Jev-like models on RAG agent routing (image 1 of 3)
Benchmarking Jev and open-source Jev-like models on RAG agent routing (image 2 of 3)
Benchmarking Jev and open-source Jev-like models on RAG agent routing (image 3 of 3)
大數軟體LargitData @LargitData1Originally posted 2026.09.20
Open original on X

Jev 一紅,社群就開源了許多 Jev-Like 模型,但到底行不行,我們實測一下! 我做了一個 RAG Agent Routing Benchmark,這次不測回答品質,而是看 Agent 能不能判斷該查知識庫、讀文件、上網、呼叫工具、追問使用者,還是直接回答。 100 題多輪對話結果: Gemma 4 31B(約束編碼):77.0% Jev:61.4% djev-spark:32.2% SemIf:24.0% Laya 322M:0% 有趣的是,只看「答案去哪裡找」,Gemma 90.0%、Jev 90.6%,幾乎一樣。真正拉開差距的是多輪狀態更新,例如「接下來只看這份文件」這種限制能不能一路維持。 Jev 雖然沒有LLM 這麼準,但優勢就是速度快,p50 749ms、p95 818ms。 開源方案中,djev-spark 已具備基本 Routing 能力,但 Scope 切換還不穩;SemIf 在權限與拒絕情境較弱;Laya 很快,p50 只有 189ms,但未微調前幾乎不能直接用。 看來快速拼湊的開源專案,還是一時半刻打不過人家要拿來賣錢的,但後勢可期!

x.com/largitdata1/status/2101679703673454669Archived 2026.09.21 10:30