Fireworks offers Ember-1 model in serverless research preview

Fireworks has made Ember-1, a model based on Kimi K3, available as a research preview on its Serverless platform. The company says it trained the model to use fewer tokens for internal reasoning while retaining its performance. Across seven benchmarks and two customers' production traffic, Fireworks reports 35-50% shorter reasoning without an accuracy loss. In live tests of the customers' coding workloads, Fireworks reports approximately 35% fewer tokens per task at comparable quality. Developers get two weeks of serverless access to the research release; permanent availability depends on community demand. Fireworks also introduced training support for companies that want to customize Ember-1 with their own data.