Reflection Beam 1.0 与 DeepSeek V4.1 Flash 在 90 道难题上对比：67 比 6

Aleksey Smolenchuk @lxe原推发布于 2026.10.20

在 X 上打开原推

Reflection @reflection_ai Beam 1.0 vs DeepSeek V4.1 Flash 1017 on 90 hard specimens (BFCL tool calls, Omni-MATH 7+, SuperGPQA hard), one sample each, code grader on @pocketmodelai: Beam 67/90
DeepSeek 66/90 Beam is 6x the cost. Maybe a lossy quant similar to my dsx experiment can help equalize the pricing here. They are similarly sized models.