Fireworks Research has introduced Ember-1, a specialized artificial intelligence model that achieves performance comparable to Kimi K3 while using approximately 40% fewer input tokens. The model was built by training on Kimi K3’s architecture and then refining its reasoning process to eliminate unnecessary computational steps without compromising output quality.
Testing across multiple dimensions confirmed the efficiency gains. Ember-1 maintained accuracy on external benchmarks, matched Kimi K3’s results in live customer A/B tests for coding tasks, and showed no measurable drop in performance during internal developer usage. In one customer deployment, the model reduced token consumption by 35% while preserving task completion rates and success metrics.
The development required more than 50 training iterations and 200 evaluation cycles. Fireworks Research created new training methods to selectively shorten reasoning chains, focusing only on the steps that directly contributed to correct answers. The process used Fireworks Serverless Training infrastructure, allowing rapid experimentation without managing dedicated GPU resources.
A key innovation addresses the disproportionate token cost of extended reasoning in multi-turn interactions. In agentic workflows, earlier reasoning traces are repeatedly reprocessed, causing context growth that scales quadratically with interaction length. Ember-1’s training specifically targeted this pattern, teaching the model to prune non-essential reflections while preserving self-correcting behaviors that help recover from errors.
Independent evaluation using the Specialized Intelligence Index benchmark showed Ember-1 establishing a new efficiency benchmark on Doximity’s Bedside Bench physician task suite. Across seven public benchmarks and two production customer workloads, the model consistently matched Kimi K3’s quality at reduced token expenditure, with token savings ranging from 35% to 50% depending on task complexity.
Fireworks Research plans to expand this approach through additional specialized models accessible via their Serverless platform. The company is also offering training support for enterprises seeking to develop custom token-efficient models using their proprietary data. Ember-1 is currently available as a research preview alongside Kimi K3, with access initially limited to two-week research periods for community testing.