10GB RTX 3080 个人助理榜：Sage-A1.5 登顶但被指循环失控
博主在 10GB RTX 3080 的「个人助理」跑分榜上把上海 AI 实验室的 Sage-A1.5 排到第一，但更新称多人报告该模型会陷入循环，需复测。
New #1 on my 10GB RTX 3080 'personal agent' board: Sage-A1.5 (35B MoE, Q3_K). [UPDATE: Technically #1, BUT - I’ve received multiple reports of this model looping to oblivion. Too many to ignore. My tests are usually conducted with reasoning off, so that would be avoiding this concern. I will test that as a follow up.] PA Core: 87.9/100, ahead of incumbents MiniCPM-V 4.7 (85.7) and Ornith 1.5 (71.2) Decode: 51.3 tok/s, about 2x higher than both. Shanghai AI Lab built it for long-horizon research, with mid-task self-checks and changing course when a plan isn't working. Turns out that's exactly what a personal agent needs when your day changes. The elephant in the room: @PrismML Ternary Bonsai 2 27B. It's still faster (1,096.7 vs 546.0 tok/s prompt) and wins still in some areas (daily organization: 84 vs 78.3). Ternary weights are a different league for speed. But Sage takes recovery (86 vs 73), scheduling (99.3 vs 88) and the overall (87.9 vs 81.7). Background of my testing and use cases in the comments👇
这条是引用着另一条写的：x.com/paulgugai/status/2113058189025640830