Fireworks AI has released Ember-1, a specialized model created by post-training Moonshot AI’s open-weight Kimi K3 to produce shorter reasoning traces while maintaining task accuracy. Unlike simply lowering the reasoning effort setting at inference time, Ember-1 is trained to eliminate redundant reasoning and unproductive loops.

  • Ember-1 delivers Kimi K3's quality with about 40% fewer tokens across evaluations.
  • It leads K3 Max on Terminal Bench 2.1 (82.0%) and DeepSWE 1.1 (75.2%), while trailing slightly on SWE-bench Verified (92.2%).
  • Production A/B tests showed output tokens falling from 49.3K to 29.9K per task with a comparable score of 0.753 versus 0.751.
  • The model is available only via the Fireworks serverless API as a Research Preview; weights and training code are not released.

This approach allows customers to access K3’s coding capabilities at a lower cost by reducing token usage without sacrificing accuracy.