Ember-1

AI model token costs are cut by nearly half for equivalent performance.

Fireworks Research has launched Ember-1, a new specialized AI model that offers the same output quality as their Kimi K3 model but uses approximately 40% fewer tokens. This efficiency gain is achieved by training Ember-1 to cut unnecessary reasoning while retaining critical thinking processes. The development involved over 50 training experiments and 200 evaluations, utilizing new training algorithms to shorten reasoning without compromising accuracy. Ember-1 has demonstrated its performance through external benchmarks, live A/B tests with customers, and internal use by Fireworks' own developers, who reported no noticeable difference in quality. The model sets a new Pareto frontier on benchmarks like Doximity’s Bedside Bench for cost per task, outperforming other leading models. Ember-1 is now available as a research preview on Fireworks' Serverless platform, marking the first in a series of specialized models aimed at reducing token costs for developers.

Ember-1's development addresses a key pain point in agentic AI: the quadratic cost increase of reasoning tokens in multi-turn interactions. By optimizing reasoning efficiency rather than simply reducing effort, Fireworks Research has created a model that maintains quality while drastically cutting token usage. This approach is validated by performance on industry benchmarks and real-world customer A/B tests, where Ember-1 matched or exceeded Kimi K3's quality at a substantially lower cost. The success of Ember-1 signals a potential shift towards more specialized, cost-efficient models that prioritize practical application and developer economics.

The introduction of Ember-1 and Fireworks Research's commitment to a series of specialized models suggest a growing market demand for AI solutions that balance advanced capabilities with affordability. Companies that rely heavily on AI for coding, agentic tasks, or extensive processing may find Ember-1 a significantly more economical option. The model's performance on the Specialized Intelligence Index, particularly its position on the Pareto frontier against leading closed-source models, indicates a competitive advancement in the specialized AI landscape. This trend could spur further innovation in model efficiency and cost optimization across the industry.