How Ember-1 Reduces Reasoning Tokens
Ember-1 is a specialized model from Fireworks Research, trained from Kimi K3. Fireworks says it learned to preserve useful reflection and planning while removing redundant reasoning, cutting tokens by about 40% at comparable answer quality. This matters particularly for multi-turn coding agents, where earlier reasoning is repeatedly included in later calls and billed again.
Ember-1’s Training and Evaluation
The company says the team ran more than 50 training experiments and over 200 evaluations across mathematics, coding, tool use, search, and software engineering. In its published tests, Ember-1 generally approached or matched Kimi K3 at its highest reasoning setting while costing less. On SWE-bench Verified, Ember-1 scored 92.2%, compared with 93.2% for K3 at maximum effort.
Fireworks also reports A/B tests on production coding traffic from two customers. Tokens fell by roughly 35% per task while downstream measures such as completion, success scores, and failure rates generally held steady or improved. The company says its own developers did not notice the switch. Still, these findings come mainly from Fireworks’ evaluations and a small number of customer tests, so they do not establish the same savings for every workload.
The Open Questions Behind the Research Preview
Ember-1 is currently offered alongside the base Kimi K3 as a Serverless Research Preview, with longer-term availability tied to community demand. Its central promise is not merely lowering a reasoning setting, but teaching the model to reason more efficiently. Whether shorter traces reliably preserve quality in long agentic workflows, rare failure cases, and unfamiliar domains still requires broader independent testing.