Fireworks Research Cuts Kimi K3 Reasoning Tokens by 40% in Ember-1

A post-trained version of the Kimi K3 model generates shorter responses during its thinking process to lower compute requirements. Fireworks Research released the model as Ember-1, which uses roughly 40% fewer tokens in reasoning than the base model while maintaining top-tier quality.

The specialized tool is 40% faster and cheaper to operate by optimizing how it consumes tokens to increase overall efficiency. Following its release, the model has trended on the Hacker News developer forum.