deepseek-v4.1IMAGE
两台 Strix Halo 本地跑 V4.1 Flash：单用户 50 tok/s，8 人共享 108 tok/s
Q2 GGUF 专家层加约 4-bit trellis 稠密层，每 token 字节数少 37%，输出逐位一致；单用户从 45 提到 50 tok/s，代码场景 70 tok/s，8 人共享从 96 提到 108 tok/s。
KireiStudios
@KireiStudios5 天前收录原推发布于 2026.10.25
KireiStudios @KireiStudios原推发布于 2026.10.25
Our DeepSeek V4.1 Flash on two small @AIatAMD Strix Halo PCs just got a lot faster: • 50 tok/s for one user (was 45), 70 tok/s on code • 108 tok/s shared across 8 users (was 96) • Same exact output, bit for bit Q2 GGUF experts + our ~4-bit trellis dense layers: 37% fewer bytes per token, same quality.
x.com/kireistudios/status/2114097639771763005采集于 2026.10.25 12:00