Event ·Multi-source

Fireworks 推出 Ember-1：基于 Kimi K3，推理 token 用量减少约 40%

Fireworks Research 发布 Ember-1，一个基于 Kimi K3 的专用模型，目标是让每个 token 发挥更大作用。该模型生成更短的推理链，token 用量约减少 40%，同时保持顶级质量。作者认为这种在帕累托前沿上压缩 token 的方向被低估了，各实验室倾向于快速发布新模型而忽视这类优化。开源前沿模型经过这类处理仍有不少可挖掘的空间。

Current read

Fireworks 拿 Kimi K3 改了个 Ember-1，推理 token 少用 40% 还不掉分，省钱党可以看看。

6 sources58° AI heatLast updated 2026/9/27 22:43:47

Evidence

6 sources

Primary sourceelvis

2026/9/27 22:43:47View source

Related source 1Cline

2026/9/27 21:58:58View source

Related source 2IT之家

2026/9/26 10:18:15View source

Related source 3Sebastian Raschka

2026/9/27 21:45:14View source

Related source 4arXiv: DeepSeek

2026/9/29 03:19:11View source

Related source 5Aravind Srinivas

2026/9/28 15:41:27View 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