Ember-1 is a new specialized model from Fireworks Research that delivers Kimi K3’s quality with 40% fewer tokens. Built on Kimi K3, it learned to cut unnecessary reasoning while preserving essential thinking. Tested on external benchmarks, live customer A/B tests, and internal coding and agent workloads, quality remained consistent.

Available today, Ember-1 launches a series of specialized models shaped by developer needs. Fireworks Research developed Ember-1 after users requested Kimi K3’s coding capabilities at lower cost, as long reasoning traces made automated coding expensive at scale. Lowering effort settings sacrificed too much quality, so the team focused on training the model to reason more efficiently.

Over 50 training experiments and 200 evaluations were conducted, leading to new training algorithms that shorten reasoning without losing accuracy. Training occurred on Fireworks Serverless Training, enabling rapid experimentation, pay-per-use pricing, and faster research-to-launch cycles. Ember-1 was evaluated across diverse tasks to ensure token savings transferred to various workloads, then validated on the Specialized Intelligence Index, public benchmarks, and live production traffic.

The model addresses inefficiency in reasoning models like Kimi K3, which spend over 90% of tokens on internal reasoning rather than answers. This becomes costly in multi-turn agentic workloads, where prior reasoning is re-read and re-billed. Ember-1 preserves useful self-reflection while eliminating excess reasoning, using feedback from tasks and environment to learn efficient reasoning.

Training spanned mathematics, coding, instruction following, conversation, search, tool use, and software engineering, including extended interactions. Results confirm maintained quality across seven benchmarks and two customers’ production traffic.