Fireworks AI has announced the release of Ember-1, a post-trained version of its Kimi K3 model designed to optimize token efficiency. This new model learns to produce shorter reasoning traces, resulting in a substantial reduction of about 40% in token usage.
In a recent production A/B test, the output tokens per task decreased from 49.3K to 29.9K, with performance scores remaining essentially unchanged. Ember-1 is currently available as an API-only Research Preview, offered at Kimi K3 pricing.
Compiled automatically by the Tech AI Newsdesk from public AI-news sources and summarised in our own words.
Frequently asked questions
What is Ember-1?
Ember-1 is a post-trained version of Fireworks AI's Kimi K3 model designed to optimize token efficiency.
How much reduction in token usage does Ember-1 achieve?
Ember-1 results in a substantial reduction of about 40% in token usage.
Is Ember-1 available for public use?
Yes, Ember-1 is currently available as an API-only Research Preview, offered at Kimi K3 pricing.
