Ember 1 tries to shorten reasoning through training

Fireworks’ Ember 1 is a research-preview model specialised from Kimi K3 to use shorter reasoning. Rather than only reducing a runtime reasoning setting, Fireworks trained it with feedback aimed at avoiding unnecessary steps. Its reported evaluations show substantial token savings, but quality varies by task, so shorter output should not be treated as a free improvement.

Fireworks AI · Sep 23, 2026

Why it matters

Agent builders can evaluate inference efficiency against actual task success.

More context

Try a fixed set of representative tasks through Fireworks’ serverless offering, measuring completion quality alongside token use. The announcement offers an initial two-week preview and says longer-term availability depends on demand.

Keep in mind

This is a research preview with temporary access terms. Results are vendor-reported, not reproduced by 1min; check current pricing and availability before committing a workflow.

Original read