The heaviest line item in a long agent run is often the model arguing with itself. Fireworks AI went straight at that cost this week with Ember-1, spun out of Moonshot’s Kimi K3 and trained to think in fewer words.
Better than nine of every ten generated tokens can vanish into internal deliberation before any answer appears.
Agent loops multiply the problem: each turn replays the earlier traces, and the same reasoning is reread and rebilled on every later call.
Fireworks tried the obvious shortcut, dialing Kimi K3’s reasoning effort down, but the quality loss was too steep. It trained for efficiency instead, keeping the useful self-reflection while cutting redundant loops.
The result came from a broad search rather than one recipe: more than 50 training experiments and over 200 evaluations, plus algorithms the company has not published.
In live A/B tests with two customers on production coding traffic, output per task fell from 49.3K tokens to 29.9K. Reasoning collapsed 71.3 percent and total tokens 39 percent, yet the task score barely moved, 0.753 against 0.751.
The scorecards are mixed and self-reported. Ember-1 edges past Kimi K3 Max on Terminal Bench 2.1 and DeepSWE 1.1, while sitting just behind on SWE-bench Verified and SWE-Interact. Fireworks also cites Doximity’s Bedside Bench, 500 cases vetted by physicians, where it says the model opened a new frontier on cost per task.
Pricing matches Kimi K3 exactly: $3 per million input tokens, $0.30 for cached input, $15 for output, and the savings come from emitting fewer tokens rather than a cheaper rate. Availability is narrow: Ember-1 runs as a research preview on the Fireworks serverless API, and its weights, training code and algorithms remain unreleased.
