WHAT NEW IN X10.28 WED

X 上的 AI 合集档案 · 22 个合集 · 58,924 条

minimax-m3.1IMAGE

13 个 OpenRouter 端点实测:7 个固定 205 输入 token 的端点 46% 回复自称 Claude

测 13 个 OpenRouter 端点发现:7 个恰好报 205 输入 token 的端点有 46% 回复提到 Claude,另 6 个仅 0.8%;205 与官方开源 chat template 吻合,temperature 0 时提到 Claude 升至 90%,且与 FP4 量化无关,作者怀疑是服务端配置所致。

Binkg
@dummerspast392 天前收录原推发布于 2026.10.25
13 个 OpenRouter 端点实测:7 个固定 205 输入 token 的端点 46% 回复自称 Claude(第 1 张,共 2 张)
13 个 OpenRouter 端点实测:7 个固定 205 输入 token 的端点 46% 回复自称 Claude(第 2 张,共 2 张)
Binkg @dummerspast39原推发布于 2026.10.25
在 X 上打开原推

MiniMax M3.1 Flash has a weird identity crisis. I tested 13 OpenRouter endpoints. 7 reported exactly 205 input tokens and mentioned Claude in 193/420 replies (46%). The other 6? Just 3/360 (0.8%). That 205-token count exactly matches MiniMax's official open-source chat template. At temperature 0, Claude mentions jumped to 90%. The strangest part is that this doesn't seem to be caused by FP4 quantization. Nebius runs FP4 and almost never mentions Claude. Parasail runs FP8 and does so in 48% of responses. The 205-token group also averages 86% on GPQA vs 88% for the other endpoints. My guess is that the serving configuration plays a major role, but the exact cause is still unclear.

x.com/dummerspast39/status/2114337785893249391采集于 2026.10.26 12:06