Fireworks Ember-1 cuts agent token use ~40% at K3's price

A post-trained Kimi K3, live now on a two-week preview: same $3/$15 rate card, ~71% shorter reasoning—cheaper agents if you count tokens per task.

Nowline SEP 30 11:00 PM banner

Top AI stories from the last hour

Top AI stories from the last hour

Copy markdown

  • Same rate card, far fewer tokens

    Ember-1 is Moonshot's open-weight Kimi K3 post-trained by Fireworks to reason more tersely, kept at K3's exact price: $3/M in, $15/M out, $0.30 cached, 1M context. Fireworks' A/B tests on customer workloads show ~71% fewer reasoning tokens and ~39% fewer total tokens per task. Your bill scales with tokens, not the sticker rate, so agent runs get cheaper with no rate-card change.

  • Where it wins, where it slips

    Terser reasoning isn't free. Ember-1 gains on Terminal-Bench 2.1 (80.9%→82.0%), DeepSWE 1.1 (66.4%→75.2%) and τ²-Bench Airline (64%→66%), but slips on SWE-bench Verified (93.2%→92.2%) and SWE-Interact (21.3%→20.0%). Shadow-test it against K3 on your own tasks before you flip production traffic.

  • Live now, but on a clock

    It's a research preview on a ~two-week serverless window on Fireworks and via OpenRouter, currently posting the best P50 throughput (~51 tok/s) and latency (~1.5s) of any K3-class provider. It may not stay serverless unless demand holds, so pilot it now rather than hard-wiring it in.