Fireworks AI has released Ember-1, a post-trained version of the Kimi K3 model. According to the company, the model is designed to generate shorter reasoning traces instead of reducing the amount of reasoning effort applied to a task.
In a production A/B test, output tokens per task dropped from 49.3K to 29.9K, a reduction of about 40%. This suggests the model is more concise in its intermediate reasoning steps while still completing the same tasks.
The reported approach differs from simply lowering reasoning effort, which would typically trade quality for speed or cost. Instead, Ember-1 learns to compress its reasoning traces, potentially offering efficiency gains without a corresponding drop in reasoning depth.