# 推理 · 日跟踪 · 2026-10-07

> 日期：2026-10-07（Asia/Shanghai，周三）。专题正文：`intel/topics/inference.md`。原料：`/workspace/intel/archive/2026-10-07/`（窗口 10-06 早报后→10-07 早）；对照 2026-10-06 日跟踪。抓取：Saved 275 / skipped 27 / errors 39；blocked_hosts：blocksandfiles.com、e.huawei.com、support.huawei.com、www.nextplatform.com、www.reddit.com、www.theregister.com（docs.redhat.com 本日未进 blocked；Virt 4.19 文档页 HTTP 503）。全库新增 URL 7 个。

## 摘要
- **中（媒体评测）· DGX Station GB300**：STH 称 Qwen3.6-35B-A3B Flash NVFP4 "well over 26,000 tokens/second"（文中 2.2B / 2.25B tokens/day 两种写法）；252GB HBM3E——**赞助评测，单机实测，非 MLPerf**。
- 续档：VMware AI Part 3（MLPerf Inference 5.1 on VCF 9.1，厂商自测；本日仅补联系人邮箱）；RH MLPerf v6.1 CPU 推理；KServe Modelcar。

## 原文摘录

### STH · Gigabyte W775-V10-L01 Hands-on Bringing NVIDIA GB300 Deskside（赞助评测）
- 归档：`industry/gigabyte-w775-v10-l01-hands-on-bringing-nvidia-gb300-deskside-d26f23f6.txt`
- URL：https://www.servethehome.com/gigabyte-w775-v10-l01-hands-on-bringing-nvidia-gb300-deskside/

> We managed to get crazy performance from this in unexpected places, like Qwen3.6-35B-A3B Flash NVFP4, a very useful model, running at well over 26,000 tokens/second or over 2.25B tokens/ day.