Fireworks 发布 Ember-1：基于 Kimi K3，token 用量省近四成

推理模型通常把 90% 以上的输出 token 花在思考上，在智能体循环里会反复思考同样的问题。Fireworks RL 用真实智能体编程任务循环做训练，教会模型区分哪些推理能改变答案、哪些只是原地打转。

在线 A/B 测试中，Ember-1 在相同成功率下推理 token 减少 71%，总 token 减少 39%。

Ember-1 by the Fireworks Research team is built on Kimi K3, and uses ~40% fewer tokens while achieving the same performance on benchmarks.

This was accomplished by post-training K3 to think less repetitively.

Reasoning models spend most of their output tokens (sometimes 90%+) on thinking before they answer, which gets expensive in agentic loops where the model tends to re-think the same thoughts on every step.

Fireworks RL trained on real agentic coding task loops to teach the model which reasoning actually changes the answer vs which is just looping.

In a live A/B test on coding traffic, Ember-1 used 71% fewer reasoning tokens and 39% fewer total tokens than K3 at the same success rate.

本条情报归属 AITOP「模型」核心栏目，重点聚焦北美前沿实验室与硅谷科技巨头。在跨区域产业观察坐标中，该动态直接关联「中美大模型技术代差与商业化对照」「AI 开发者工具链、Coding Agent 与工程化实践」研究主线，反映出当前技术演进与商业落地的最新风向。涉及核心实体与技术议题包括：Ember-1、Fireworks、Kimi K3、推理模型，读者可通过下方标签进入对应专题页查阅历史脉络与关联对比。

信源核验说明：本文原始线索采编自「Cline」（@cline），经 AITOP 多维评分管线结构化提炼。截至目前，全网已有 6 家独立信源对该事件进行交叉印证与跟进报道（含 elvis、IT之家、Sebastian Raschka 等），属于近期高关注度行业信号。