Event ·Official first-hand

80张RTX 5090运行Kimi K3实现20 tok/s

开源模型Kimi K3（2.8T参数）在80张RTX 5090上以单流20 tok/s运行。该配置使用GDDR7游戏显卡和普通以太网，无需HBM，权重为官方MXFP4格式。此前同一集群运行GLM-5.2时从30 tok/s优化至110 tok/s。任何实验室或大学均可使用该配置复现。

Current read

80张游戏卡就能跑2.8T参数的Kimi K3，20 tok/s挺实用，比之前GLM-5.2快了好几倍。

11 sources62° AI heatLast updated 2026/7/28 11:12:46

Evidence

11 sources

Primary source歸藏(guizang.ai)

2026/7/28 11:12:46View source

Related source 1SuperTechFans

2026/7/28 00:12:40View source

Related source 2IT之家

2026/7/28 08:17:30View source

Related source 3Yangyi

2026/7/27 15:26:08View source

Related source 4AK

2026/7/27 15:29:00View source

Related source 5Fireworks AI

2026/7/27 15:30:35View source

Related source 6Cognition

2026/7/27 15:38:20View source

Related source 7Guillermo Rauch

2026/7/27 15:38:25View source

Related source 8Windsurf

2026/7/27 15:39:43View source

Related source 9The Rundown AI

2026/7/27 15:39:52View source

Related source 10elvis

2026/7/27 15:42:57View source

Conflict check

Dedup data only shows that these sources discuss the same event, not that they agree. There is no structured supporting or refuting evidence yet, so no conflict is inferred automatically.

Open the story