Event ·Multi-source

inferact 开源 Kimi K3 TPU megakernel，单核跑出 709 tokens/s

inferact 团队为 Kimi K3 开源了一个 TPU megakernel，推理速度达到 709 tokens/s，高于 GB200 上的 450 tokens/s。K3 全部 92 个 MoE 层都在单个 Pallas kernel 中运行。该实现采用跨层边界的权重预取，让一层的权重传输与上一层的计算重叠进行。代码仓库和详细写文已发布。

Current read

Kimi K3 在 TPU 上跑到 709 tokens/s，比 GB200 还快，92 个 MoE 层塞进一个 Pallas kernel，做法挺巧，做推理的可以看看。

7 sources58° AI heatLast updated 2026/9/23 18:20:22

Evidence

7 sources

Primary sourcevLLM

2026/9/23 18:20:22View source

Related source 1Thomas Wolf

2026/9/21 19:50:48View source

Related source 2eric zakariasson

2026/9/21 22:01:00View source

Related source 3IT之家

2026/9/21 23:40:31View source

Related source 4Fireworks AI

2026/9/23 21:38:14View source

Related source 5pandaily

2026/9/23 01:53:48View source

Related source 6arXiv cs.AI

2026/9/24 12:23:52View 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