Fireworks Releases Ember-1: Matches Kimi K3's Performance with 40% Fewer Tokens

Fireworks Research has introduced Ember-1, a new specialized model designed to reduce the token usage of reasoning models while maintaining high quality. Built on Kimi K3, Ember-1 achieves a 40% reduction in token consumption by optimizing reasoning traces, specifically targeting unnecessary steps while preserving critical self-reflection capabilities.

The model was evaluated on several industry benchmarks, establishing a new Pareto frontier for cost-efficiency. On Terminal Bench 2.1, Ember-1 achieved an 82.0% pass rate compared to 80.9% for Kimi K3, and on SWE-bench, it scored 92.2% versus 93.2% for K3, all while using significantly fewer tokens. The model also outperformed closed models like GPT-6 Astra and Claude Opus 5 on the Bedside Bench in terms of cost-per-task.

In live A/B tests with two customers, Ember-1 demonstrated a 39% reduction in total token usage with comparable or improved task completion rates. The model is currently available as a Research Preview on Fireworks Serverless, with training support announced for enterprises looking to customize the model for specific workloads.