Fireworks AI unveils Ember-1, a Kimi K3-based model that cuts tokens 35%

Fireworks AI describes its new Ember-1 as the first release from a project to build the specialized models developers need, targeted at a specific complaint: that cutting-edge models spend so much of their output on reasoning that costs balloon, especially when an AI agent reloads the whole context on every turn.

The approach was to add training to Kimi K3, keeping the parts of its reasoning that help complete the task, such as feedback, while trimming the rest. On the Bedside Bench medical task benchmark, Ember-1 held Kimi K3's score at a lower cost, and its processing time per task was shorter. Kimi K3 had been seen as costing more while performing below GPT-5.6 Sol; Ember-1 sits as the cheaper option that still trails GPT-5.6 Sol on performance.

What hinges on the outcome is whether the roughly 35% token reduction, confirmed in an A/B test with two customers, holds beyond those tests and for Fireworks AI's own staff. For developers and companies paying per token to run agents, that gap between price and performance is the number that decides whether they switch.