Event ·Multi-source

vLLM 上线 Recursive SI 的 Flint Day-0 支持：40B 参数仅 2.5B 激活

vLLM 宣布对 Recursive SI 的 Flint 提供 Day-0 支持。Flint 共 40B 参数，每个 token 仅激活 2.5B。借助 vLLM 的自定义 NVFP4 内核，单张 GB200 在 100K 上下文、不开投机解码的情况下达到 500 TPS。NVFP4 和 BF16 两种精度目前均可直接运行。

Current read

vLLM 第一天就支持了 Recursive SI 的 Flint，40B 混合专家模型，单张 GB200 跑出 500 TPS，还支持 100K 长上下文。

8 sources58° AI heatLast updated 2026/10/21 16:45:06

Evidence

8 sources

Primary sourcevLLM

2026/10/21 16:45:06View source

Related source 1rohanpaul_ai

2026/10/21 18:21:28View source

Related source 2marktechpost

2026/10/21 22:39:41View source

Related source 3Teortaxes

2026/10/20 23:47:42View source

Related source 4IT之家

2026/10/21 15:25:18View source

Related source 5Richard Socher

2026/10/21 16:29:32View source

Related source 6TestingCatalog

2026/10/21 19:06:27View source

Related source 7techcrunch

2026/10/21 20:27:41View 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