deepseek-v4.1POST
3/4 台 DGX Spark 本地跑 V4.1 Flash：4 卡 87 tok/s
用 DeepSeek 原版量化权重在 DGX Spark 上部署：4 卡单流 87 tok/s、4 并发 163 tok/s、prefill 5000 tok/s；3 卡单流 51 tok/s、prefill 2000 tok/s。
Mia
@MiaAI_lab11 天前收录原推发布于 2026.10.18
Mia @MiaAI_lab原推发布于 2026.10.18
Run DeepSeek V4.1 Flash 1037 on 3/4 DGX Sparks 🔥 This uses the original quant posted by DeepSeek, on both 3 and 4x Sparks. No further quantization was made. 4x DGX Sparks: - 87 tok/s on prose, single stream - 163 tok/s on prose, 4 concurrent streams - Prefill 5000 tok/s across the board. 3x DGX Sparks: - 51 tok/s on prose, single stream - 163 tok/s on prose, 4 concurrent streams - Prefill 2000 tok/s across the board. Also merged a lot of PRs and fixed a lot of issues. Get it here: https://mia-ai.net/models/DeepSeek-v4.1-Flash-DGX-Sparks
x.com/miaai_lab/status/2111554301424230681采集于 2026.10.18 16:28