Fireworks AI has released Ember-1, a specialized model built by post-training Moonshot AI’s open-weight Kimi K3 to use about 40 percent fewer tokens while maintaining task accuracy. Developed by the Fireworks Research team, Ember-1 learns to produce shorter reasoning traces rather than simply lowering the reasoning effort setting at inference time. The model is currently available as a Research Preview through the Fireworks serverless API, though self-hosting is not supported since the weights and training code have not been released.
Reasoning models often spend more than 90 percent of their generated tokens on internal reasoning, which drives up costs in multi-turn agentic workloads where context grows quadratically. To solve this, Fireworks trained Ember-1 on a diverse collection spanning mathematics, coding, instruction following, and tool use, utilizing task and environment feedback to cut redundant reasoning loops while preserving useful self-reflection. In live production A/B tests with coding customers, Ember-1 reduced total tokens by 39 percent and reasoning tokens by over 71 percent while maintaining comparable task scores.
This release matters because it lowers the operational cost of advanced reasoning models for enterprise applications without sacrificing performance.