Ember-1

Ember-1, built on Kimi K3, reduces token usage by 40% while maintaining the same quality on coding, mathematics, instruction-following, conversation, search, tool use, and software-engineering tasks. The model learns to trim unnecessary internal reasoning, cutting 35–50% of reasoning tokens without accuracy loss, as shown on seven public benchmarks and two customers’ production traffic. Ember-1 was trained with over 50 experiments and 200 evaluations on Fireworks Serverless Training, enabling rapid deployment and cost savings.