OpenUnum 用 Jev 做 System One 决策层

atonioz.ΕΘ @at0nios原推发布于 2026.09.27

在 X 上打开原推

I gave OpenUnum a "System One": a fast layer that makes the agent's small, typed decisions instead of asking the big model every time. Four decisions, every turn: 1. Is the answer actually backed by the tool results? 2. Did the subagent really finish? 3. Should this task be delegated? 4. Does this request need the shell? Scored TypeSafe's Jev against the open "Jev lookalike" models, all run on one Ryzen AI MAX+ 395 box, with the open ones sandboxed in containers: • Jev: 100 / 100 / 100 / 95%, about 0.3 s per decision • CLM-8B: 95% on tool routing at 87 ms, matching Jev and 3–4× faster, but about 50–57% on the yes/no checks • Laya, Von, Kev, Verdict, Julia-1, poorjev: 47–67% on grounding • The best general LLMs reach 80–97% but take 0.4–5 s What OpenUnum runs now: • Tool routing: CLM-8B, self-hosted, with Jev as fallback • The other three: Jev, with local fallbacks (main model, built-in rules) • Every decision has a 4 s cap and quietly falls back Small hand-labelled test sets, so this shows who can do the job, not production accuracy. Scoreboard below. #amd #bosgame #strixhalo #OpenUnum #SystemOne #Jev #TypeSafe #CLM #ContrastiveLM #Laya #Julia1 #Von #Kev #Verdict #poorjev #Qwen #Gemma #Nemotron #gptoss #llamacpp #Podman #LocalAI #AIAgents