Fireworks AI launches Ember-1, saying token usage drops about 40% versus Kimi K3

Fireworks AI has introduced Ember-1, an inference-optimized model built through post-training on Moonshot AI’s open-weight Kimi K3. The company said the model is designed to preserve task accuracy while producing shorter reasoning traces, cutting token usage by about 40% compared with the original Kimi K3. Ember-1 is currently available only as a research preview through Fireworks’ serverless API. Its weights, training code, and specific algorithms have not been released, and self-hosting is not supported at this stage. Fireworks said the goal is to keep K3’s coding capability while lowering customer costs. In internal testing, the company found that more than 90% of tokens generated by K3 in multi-turn agent workloads were spent on internal reasoning, creating significant cost overhead. In production A/B tests covering programming workloads from two customers, Fireworks said Ember-1 reduced tokens per task by about 35% at similar quality levels. In one test, output tokens fell from 49,300 to 29,900 per task, reasoning tokens dropped 71.3%, total tokens declined 39%, and task scores were broadly unchanged.