RC RANDOM CHAOS

It walks through the handshake between your apps

Fireworks trained Ember-1 to cut Kimi K3's reasoning tokens by 35 to 50% without losing accuracy, aimed at agentic coding where reasoning is the cost.

· 3 min read
It walks through the handshake between your apps

A reasoning model like Kimi K3 spends most of its output tokens on internal reasoning rather than on the answer it returns. Fireworks puts that share at sometimes more than 90%. On a single request that is already expensive. In a multi-turn agentic loop it compounds, because every turn replays all prior reasoning back to the model. Context grows roughly quadratically with the number of turns, and the long reasoning traces from early turns get re-read and re-billed on every later call.

Fireworks Research says it trained a version of K3 that emits less of that reasoning without losing accuracy. The model is Ember-1, built on Kimi K3, sold as “same answers, fewer tokens.” The headline claim is about 40% fewer tokens. Across seven benchmarks and two customers’ production traffic, Fireworks says K3’s reasoning could be shortened by 35 to 50% with no drop in accuracy.

The obvious knob came first. Fireworks tried turning K3’s reasoning effort down and says the lower settings gave up too much quality, so it trained the model instead. The training collection spanned mathematics, coding, instruction following, conversation, search, tool use, and software engineering, covering both standalone problems and extended interactions. Task feedback guided on-policy learning, the idea being to connect decisions to their consequences so the model learns which reasoning actually pays off. Fireworks reports more than 50 training experiments and over 200 evaluations, run on its own Serverless Training, using its own data and no customer data.

The useful nuance is that not all of K3’s reasoning is padding. Some of it is self-reflection, revisiting an assumption or tracing an outcome back to an earlier decision, and that helps the model recover from mistakes. The goal was to keep that behaviour while cutting the excess and escaping unproductive loops. Fireworks says the trained model also shows restrained token use on attempts it fails, so it does not spiral on problems it cannot solve.

What Fireworks measured

The numbers are self-reported, the A/B tests are the vendor’s, and the cost figures are computed by the vendor, so read them as claims.

On Doximity’s Bedside Bench, a physician-validated benchmark of 500 clinical cases across 10 categories, Ember-1 set a new Pareto frontier on cost per task against GPT-5.6 Sol, GPT-6 Astra, and Claude Opus 5. On cost versus quality across benchmarks with more than 50 samples, priced against the public K3 API ($3 per million uncached input tokens, $0.30 cached, $15 output), Fireworks says Ember-1 sits on or near the Pareto frontier: matching K3 at its default max reasoning effort for a fraction of the cost, and strictly beating K3 at low effort. The takeaway Fireworks draws is that the cheapest way to run K3 is now Ember-1 rather than a lower reasoning-effort setting.

Two customers ran A/B tests on production coding workloads and saw roughly 35% fewer tokens per task at comparable quality, with task completion, success scores, and failure rates holding or improving. One is now running Ember-1 in production and plans to replace the base model. Fireworks also ran it on its own internal coding traffic first, and reports that developers did not notice the switch while spending fewer tokens.

If you run agentic workloads

The structural point holds whatever you make of the benchmarks: in a multi-turn agentic system the reasoning trace is the cost, and that cost is quadratic in turns. If reasoning dominates your bill, the reasoning-effort setting is a blunt instrument, because it trades quality for tokens linearly. A model trained to spend tokens where they matter is a different lever. Before reaching for either, measure where your tokens actually go per turn, since the quadratic replay means early-turn verbosity is what you pay for most.

Ember-1 is out as a Research Preview on Fireworks Serverless, serving alongside base K3. Research releases get two weeks of serverless access and become permanent based on demand. Fireworks is also launching training support for Ember-1, so you can fine-tune a token-efficient model on your own workload data.

Share

Keep Reading

Latest on the Wire

Full wire →

New signal daily · RSS

Stay in the loop

New writing delivered when it's ready. No schedule, no spam.