Ember-1
Points and comments are a snapshot, not live.
Fireworks' Ember-1 matches Kimi K3 quality with ~40% fewer tokens by trimming unnecessary reasoning.
Fireworks Research built Ember-1 by training Kimi K3 to reason more efficiently, cutting 35-50% of tokens across benchmarks without accuracy loss. It uses the Fireworks Serverless Training platform and no customer data. On Doximity's Bedside Bench, Ember-1 set a new Pareto frontier against GPT-5.6 Sol, GPT-6 Astra, and Claude Opus 5. Live A/B tests with two customers showed ~35% token savings with comparable quality. Internal developers did not notice the switch. Ember-1 is available as a Research Preview on Serverless, with training support for enterprise customization.
What commenters are saying
Commenters focused on the model's limited openness: trained on open weights but weights not released. Several noted the term 'Pareto frontier' is overused. Some questioned what specific capabilities were lost, while others found the techniques likely generalizable to larger models. Skepticism emerged about opaque training descriptions versus actual technical detail. Comparisons to GLM 5.3 Flash were raised. One camp argued open models can still advance rapidly through community iteration.