Fireworks AI Releases Ember-1: A Post-Trained Kimi K3 That Uses About 40% Fewer Tokens
Fireworks AI launched Ember-1, a post-trained Kimi K3 that cuts reasoning tokens ~40% at matched accuracy, API-only.
Fireworks AI released Ember-1, a model built by post-training Moonshot AI's open-weight Kimi K3 to produce shorter reasoning traces, claiming Kimi K3 quality with roughly 40% fewer tokens. It beats K3 Max on Terminal Bench 2.1 (82.0% vs 80.9%) and DeepSWE 1.1 (75.2% vs 66.4%), trails slightly on SWE-bench Verified (92.2% vs 93.2%). In production A/B tests with two customers, output tokens fell from 49.3K to 29.9K per task with scores essentially unchanged (0.753 vs 0.751). Ember-1 is available only through the Fireworks serverless API as a Research Preview at Kimi K3 pricing ($3.00 input / $15.00 output per 1M tokens); weights and training code are not released.
- Ember-1 post-trains Kimi K3 to reason in ~35-50% fewer tokens without sacrificing accuracy.
- Leads K3 Max on Terminal Bench 2.1 (82.0%) and DeepSWE 1.1 (75.2%), trails on SWE-bench Verified (92.2%).
- Production A/B: 49.3K to 29.9K output tokens per task, score 0.753 vs 0.751.
- API-only Research Preview; weights, training code and algorithms not released.
- Organizations
- Fireworks AIMoonshot AIEmber-1
- AI models
- Kimi K3
From the original article
4 min read at marktechpost.com
Fireworks AI has released Ember-1, a post-trained Kimi K3 that learns to produce shorter reasoning traces instead of lowering reasoning effort. Fireworks reports about 40% fewer tokens, with output tokens per task falling from 49. 3K to 29.
By Asif Razzaq · marktechpost.com
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.