Ember-1

Ember-1 is a Fireworks Research specialized reasoning model built on Moonshot's Kimi K3 that shortens reasoning traces by roughly 40% while matching K3-max quality on Fireworks' internal evaluations. It is a mixture-of-experts model with 2.78T parameters, a 1.04M-token context, and function calling, offered as a Research Preview on Fireworks Serverless under the model path `accounts/fireworks/models/ember-1`. List pricing is $3.00 input / $0.30 cached input / $15.00 output per million tokens, the same as public Kimi K3 API pricing. The four catalog scores are self-reported from the Fireworks launch blog (Terminal-Bench 2.1 82.0%, SWE-bench Verified 92.2%, DeepSWE 1.1 75.2%, tau2-bench Airline 66%).