Fireworks AI released Ember-1 today, a compressed version of its Kimi K3 model that the company says holds K3’s answer quality while spending 40 percent fewer tokens per task. The launch runs as a two-week Research Preview on Fireworks Serverless. Fireworks will decide later, based on developer demand, whether to keep it available permanently.
Token spend, not a new capability, is the actual story. Agentic coding tools built on long reasoning chains pay for every internal token the model generates, and that bill grows worse across multi-turn tasks: each new call re-sends the full prior reasoning trace back into context, so the cost compounds turn over turn. Cutting the trace without cutting quality is a cost problem wearing a research label, and it is the same pressure nudging labs like DeepSeek toward distillation and Anthropic toward tiered pricing.
Fireworks says the work behind Ember-1 involved a large research push. Its team tried more than fifty training setups and logged upwards of two hundred evaluation runs, building new training methods so K3’s reasoning could shrink without losing the self-correction that keeps answers accurate. The company reports that across seven internal benchmarks, plus production traffic from two customers, the shortened reasoning held accuracy steady while cutting length by 35 to 50 percent.
On Doximity’s Bedside Bench, a physician-validated clinical test, Fireworks reports that Ember-1 sits on a new Pareto frontier for cost per task, ahead of figures the company attributes to GPT-5.6 Sol, GPT-6 Astra, and Claude Opus 5. Those numbers come from Fireworks’ own infrastructure, run by Fireworks itself, not from an independent evaluator.
The company also published a head-to-head against unmodified K3 running at its highest reasoning setting. Fireworks says Ember-1 scored 92.2 percent on SWE-bench Verified against K3’s 93.2 percent, trading a small accuracy gap for a claimed 15.5 percent cost cut. On a different test, DeepSWE 1.1, the company reports Ember-1 ahead of K3 outright, 75.2 percent versus 66.4 percent, while still spending less. Across the five benchmarks Fireworks disclosed, the pattern reads as close to even on raw accuracy and consistently cheaper on tokens, not a win on every measure.
Two customers also ran live A/B tests on production coding workloads. Fireworks says Ember-1 held quality while cutting token use by roughly 35 percent per task, with completion and failure rates holding steady or improving. One of those customers has since moved the model into full production and, according to Fireworks, intends to swap it in for the base model across the board. Fireworks frames its own internal deployment, where staff kept coding without noticing which model they were running, as its clearest evidence, even though that is also the claim with the least outside verification.
None of the benchmark figures or the A/B results have been checked by an outside party, and Fireworks has not named the two customers or their industries. The company is also opening training access so other businesses can build their own token-efficient models on their own data, turning Ember-1’s cost story into a lead-in for the Fireworks Training platform.
For teams already running agentic coding workflows on Kimi K3 through Fireworks, the two-week window is tight. Any migration call needs its own benchmark pass against real workloads before the preview ends or turns into a paid tier.
Fireworks AI detailed the Ember-1 release on its own company blog, published the day the model went live.
