Fireworks AI releases Ember-1, a token-efficient version of Kimi K3

Fireworks AI has released Ember-1, an open-weight model whose underlying code and parameters anyone can download, built on top of Kimi K3, according to a blog post from Fireworks Research. The company says Ember-1 matches Kimi K3’s quality while using far fewer tokens, the chunks of text a model reads and writes, to get there.

The problem Fireworks says it was solving is cost. Reasoning models like Kimi K3 generate long internal reasoning traces, draft steps the model writes before its final answer, and Fireworks says these traces can account for more than 90% of a response’s tokens. That cost compounds in multi-turn agentic tasks, where a model calls tools or takes actions across several steps, because each new turn re-reads all the prior reasoning. Fireworks says simply lowering Kimi K3’s reasoning effort setting cut quality too much, so it trained a new model instead.

According to Fireworks, the company ran more than 50 training experiments and 200 evaluations on its own Serverless Training platform, using only Fireworks’ own data rather than customer data. The training set spanned math, coding, instruction following, conversation, search, tool use and software engineering tasks, aimed at teaching the model which reasoning steps to keep, such as catching its own mistakes, and which to drop.

Fireworks says Ember-1 shortened Kimi K3’s reasoning by 35 to 50% across seven public benchmarks and two customers’ production traffic without a drop in measured accuracy. On Doximity’s Bedside Bench, a physician-validated benchmark covering 500 clinical cases, Fireworks says Ember-1 set a new cost-to-quality result against both open and closed models. In live A/B tests on production coding workloads with two customers, the company reports roughly 35% fewer tokens per task at comparable quality, with task-completion and failure-rate metrics holding steady or improving; one of those customers has since moved Ember-1 into full production. Fireworks also says its own developers used Ember-1 for everyday coding work without noticing a difference from Kimi K3, using substantially fewer tokens along the way.

Ember-1 is rolling out today as a two-week Research Preview on Fireworks Serverless, running alongside the base Kimi K3 model. Fireworks says access will become permanent based on developer demand, and it is also opening training support so businesses can fine-tune Ember-1 on their own workloads. What remains to be seen is how the model performs under independent testing outside Fireworks’ own benchmarks and customer trials, and whether Fireworks follows through on its plan for further specialized models in the Ember series.