Fireworks AI has introduced Ember-1, a post-trained version of the Kimi K3 model. The new model is currently available exclusively through the Fireworks serverless API as a Research Preview. The company has not released the model weights, training code, or specific algorithms, making self-hosting unavailable.
The development of Ember-1 focused on reducing redundant reasoning tokens. Fireworks reports that Kimi K3 often spends over 90% of tokens on internal reasoning, which increases costs in multi-turn agentic tasks. Ember-1 maintains essential self-reflection while eliminating unproductive loops, resulting in a 35% to 50% reduction in reasoning tokens across seven benchmarks.
In production A/B tests with two customers, Ember-1 demonstrated a 39% reduction in total tokens per task while maintaining comparable quality. Specifically, output tokens decreased from 49.3K to 29.9K, and reasoning tokens dropped by 71.3%. The task score remained stable at 0.753 compared to 0.751 for Kimi K3, with average steps per task falling from 23.8 to 21.4.
The training process for Ember-1 involved over 50 experiments and 200 evaluations using the Fireworks Serverless Training platform. The company utilized its own proprietary data for training, excluding any customer data. The model showed strong performance on benchmarks such as Terminal Bench 2.1 and DeepSWE 1.1, and established a new cost-per-task frontier on the Doximity Bedside Bench.
Pricing for Ember-1 is identical to Kimi K3 on the Fireworks platform, set at $3.00 per 1M input tokens, $0.30 per 1M cached input tokens, and $15.00 per 1M output tokens. Cost savings for users are achieved solely through the reduction in the total number of tokens generated during task execution.
