Fireworks AI unveils ember-1: a more efficient version of kimi k3 with 40% fewer tokens

Fireworks AI has introduced Ember-1, a post-trained variant of Moonshot AI’s Kimi K3, designed to generate about 40% fewer tokens while maintaining task accuracy. The model is available via Fireworks’ serverless API as a Research Preview, though self-hosting is not currently possible due to the lack of released weights, training code, or algorithms. Fireworks claims that Ember-1 reduces redundant reasoning without sacrificing quality, addressing the high token costs associated with multi-turn agentic workloads. The model was trained on a diverse dataset covering mathematics, coding, and software engineering, with over 50 experiments and 200 evaluations. Fireworks reported that Ember-1 outperformed Kimi K3 Max on Terminal Bench 2.1 and DeepSWE 1.1, while slightly trailing on SWE-bench Verified and SWE-Interact. In live A/B tests with two customers, Ember-1 reduced tokens per task by 35-50% without affecting task scores. One customer has already deployed Ember-1 in production. The model costs the same per token as Kimi K3 on Fireworks’ platform, with savings coming from fewer token generation.