Ember-1
5 hours ago
- Fireworks Research launched Ember-1, a specialized model delivering Kimi K3 quality with 40% fewer tokens by learning to cut unnecessary reasoning.
- The model was trained through over 50 experiments and new algorithms, using Fireworks Serverless Training, and tested on benchmarks, live A/B tests, and internal coding workloads.
- Reasoning models like Kimi K3 spend over 90% of tokens on internal reasoning; Ember-1 preserves essential self-reflection while eliminating excess, achieving 35-50% token reduction without accuracy loss.
- On the Specialized Intelligence Index, Ember-1 set a new Pareto frontier on Doximity's Bedside Bench, outperforming models like GPT-5.6 Sol and Claude Opus 5 on cost per task.
- Live A/B tests with customers showed ~35% token savings per task with comparable or improved quality metrics, leading to production adoption.
- Internal rollout at Fireworks was seamless—developers noticed no quality change while consuming substantially fewer tokens.
- Ember-1 is available as a research preview, with plans for more specialized models and training support for enterprise customization.