deepseek-v4.1IMAGE
不限思考时间跑分：V4.1 Flash 87%，实时全 0%
给足思考时间时 DeepSeek V4.1 Flash 得 87%，GPT-6 Astra 74%、Muse 62%；但切到实时模式所有大模型都掉到 0%，微调过的 Qwen 4B 反而有 90%。
Zhanghao Wu
@Michaelvll19 天前收录原推发布于 2026.10.23
Zhanghao Wu @Michaelvll1原推发布于 2026.10.23
With unlimited time to think: DeepSeek V4.1 Flash 87%, GPT-6 Astra 74%, Muse 62%, Nemotron 60%, Beam 58%. In real time: all 0%. The post-trained Qwen 4B model, in real time: 90%.
x.com/michaelvll1/status/2113317480970277340采集于 2026.10.23 12:07