Fireworks trained a model to think 40% less for the same answers.
Fireworks Research retrained Moonshot's open Kimi K3 model to stop it explaining itself at such length. The result, Ember-1, matched K3 across seven benchmarks while spending far fewer thinking tokens. It is a two-week research preview, so it survives only if enough people use it.
Why it mattersReasoning models bill you for every word they think, and much of that thinking is filler. Cheaper thinking is what moves an agent from a demo to something you can afford.
Read in the app — free, in 9 languages
Related stories
Amazon now sells a way to watch what your AI agent actually did.2026-09-28Google says Gemini 4 has entered its last stage before release.
2026-09-28A startup three months old beat OpenAI on a robot benchmark.
2026-09-27Meituan's 1.6-trillion-parameter model learned to read screenshots.
2026-09-26The creator of Rails writes code only when an agent gets it wrong.
2026-09-26