Ember-1 is a specialized reasoning model from Fireworks Research built on Moonshot AI's Kimi K3. Post-training shortens reasoning traces while preserving task accuracy, yielding 35-52% token reductions on coding and agent benchmarks such as Terminal Bench, SWE-bench Verified, and DeepSWE. It supports 1M context, image input, tool use, and ships on Fireworks serverless at $3/$15 per million tokens.
