Fireworks AI releases Ember-1 for shorter Kimi K3 reasoning
Fireworks AI has introduced Ember-1, a specialized model created by post-training Moonshot AI’s open-weight Kimi K3. The model is designed to preserve useful reflection and feedback-driven reasoning while removing repetitive traces and unproductive loops. Fireworks says this approach differs from simply reducing Kimi K3’s reasoning-effort setting, which lowered quality too substantially in its tests.
According to Fireworks’ evaluations, Ember-1 uses about 40% fewer tokens while maintaining comparable task accuracy. It exceeded K3 Max on Terminal Bench 2.1, scoring 82.0% versus 80.9%, and on DeepSWE 1.1, scoring 75.2% versus 66.4%. It was slightly behind K3 Max on SWE-bench Verified and SWE-Interact. Fireworks also reports 35–50% shorter reasoning across seven benchmarks and two customers’ production traffic.
In a production coding A/B test, output tokens fell from 49.3K to 29.9K per task, while the task score remained nearly unchanged at 0.753 versus 0.751. Ember-1 is available only through Fireworks’ serverless API as a Research Preview. Its weights, training code, and exact algorithms have not been released. Pricing matches Kimi K3: $3.00 input, $0.30 cached input, and $15.00 output per 1M tokens.
Written by Sakhanda Wire AI from reporting by MarkTechPost. Facts, quotes and full context are in the original article below.