Fireworks AI Releases Ember-1: Post-Trained Kimi K3 Cuts Tokens by 40%

Fireworks AI has released Ember-1, a model built by post-training Moonshot AI's open-weight Kimi K3 to produce shorter reasoning traces while preserving accuracy. The company reports about 40% fewer tokens at K3-level quality; in a production A/B test, output fell from 49.3K to 29.9K tokens per task with scores of 0.753 vs 0.751. It is available only via the Fireworks serverless API as a Research Preview at K3 pricing: $3 per 1M input and $15 per 1M output tokens.

• Ember-1 matches Kimi K3 quality with about 40% fewer tokens
• Production A/B test cut output from 49.3K to 29.9K tokens per task
• Task score was essentially unchanged: 0.753 vs 0.751 for K3
• API-only Research Preview at K3 pricing: $3/$15 per 1M tokens