Fireworks says Ember‑1 cuts reasoning tokens 40%
Fireworks Research built Ember-1 on Moonshot AI's Kimi K3 to reason in fewer tokens, and the release topped Hacker News with 263 points hours after it shipped.
A Hacker News post about Fireworks Research's Ember-1 model reached 263 points and 145 comments Sunday, five days after the company released it. Fireworks built Ember-1 by post-training Moonshot AI's Kimi K3 to produce shorter reasoning traces. The company said the model uses roughly 40% fewer tokens while matching K3's quality.
Fireworks said Ember-1 scored 82.0% on Terminal Bench 2.1 and 75.2% on DeepSWE 1.1. Kimi K3 scored 80.9% and 66.4% on the same two benchmarks, the company said. Fireworks said it ran more than 50 training experiments and 200 evaluations, using new training algorithms built on its own serverless training platform.
Ember-1 is available now through Fireworks' serverless API as a free research preview for two weeks, the company said. After that it plans to charge $3 per million input tokens and $15 per million output tokens. The model carries a 1.04-million-token context window and supports function calling and image input, its model card says.
The token-count and benchmark figures are Fireworks' own, measured on evaluations it selected. They have not been independently verified. Dmytro Dzhulgakov, the company's co-founder and chief technology officer, said on X that the HN reception caught his team off guard, writing that Ember-1 was "hot on HN, our research team cooked."
https://x.com/dzhulgakov/status/2104311573640855903