Fireworks Research released Ember-1, a specialized model. It matches Kimi K3’s quality with 40% fewer tokens. The model learned to cut unnecessary reasoning while keeping accuracy. It was trained on Fireworks Serverless Training. Ember-1 set a new Pareto frontier on Bedside Bench. It also led on five other industry benchmarks. Live A/B tests showed 35% fewer tokens at comparable quality. Internal developers did not notice the switch. Ember-1 is available as a Research Preview on Serverless. Fireworks plans more specialized models in the Ember series.
What changed
Kimi K3 used full-length reasoning traces for all tasks.
What it unlocks
Running the same workloads at roughly half the token cost without quality loss.
- 40% fewer tokens
- 35-50% token savings in benchmarks
- 39% total token reduction in A/B tests
- 71.3% reasoning token reduction
What you need to act on it
- access to Fireworks Serverless
- Research Preview release
Sources