Ember-1 (fireworks.ai)

Fireworks Research has announced Ember-1, a new specialized model that delivers the capabilities of its predecessor Kimi K3 while utilizing 40% fewer tokens. This innovation is significant for the AI/ML community as it addresses the high token costs associated with extensive reasoning in models, particularly in coding and multi-turn interactions. Ember-1’s development involved more than 50 training experiments and the creation of new algorithms that enhance reasoning efficiency without sacrificing quality. Tests across various benchmarks demonstrated that Ember-1 not only maintains the performance of Kimi K3 but does so at a substantially reduced cost. The model excels in diverse tasks, from coding to software engineering, and has been validated through live A/B tests, showing approximately 35% fewer tokens used per task while preserving completion and success rates. Ember-1 is available for developers as a Research Preview release, aiming to foster further innovation in token-efficient AI applications and support the open-source community. With plans to introduce more specialized models, Fireworks emphasizes a shift toward developing tailored solutions that meet specific business needs, paving the way for more economical AI deployments.