Fireworks AI has released Ember-1, a specialized model from Fireworks Research built by post-training Moonshot AI’s open-weight Kimi K3. Ember-1 learns to produce shorter reasoning traces while keeping task accuracy. This is different from lowering the reasoning effort setting at inference time. According to the Fireworks release post, Ember-1 delivers Kimi K3’s quality with about 40% fewer tokens.
Is it deployable? Yes, but only through the Fireworks serverless API as a Research Preview. Fireworks has not released Ember-1’s weights, training code, or exact training algorithms, so self-hosting is not an option today.
The Problem: Reasoning Models Think Too Much
Fireworks team reports that reasoning models like Kimi K3 sometimes spend more than 90% of generated tokens on internal reasoning. That cost compounds in multi-turn agentic workloads. Each turn replays prior reasoning back to the model. Context grows roughly quadratically with the number of turns. Long traces from early turns get re-read, and re-billed, on every later call.
Fireworks team explains how customers wanted K3’s coding capability at lower cost. Turning down K3’s reasoning effort did not solve it. Lower effort settings gave up too much quality. So the team trained the model to reason more efficiently instead.
How Fireworks Research Built Ember-1
Not all of K3’s reasoning is waste. Some of it is useful self-reflection, like revisiting an assumption or reacting to feedback. Ember-1 keep that behavior while cutting redundant reasoning and unproductive loops.
The training collection spans mathematics, coding, instruction following, conversation, search, tool use, and software engineering. It covers both standalone problems and extended multi-step interactions. Task and environment feedback guides on-policy planning and learning. Fireworks team ran more than 50 training experiments and over 200 evaluations. They also developed new training algorithms, which it has not published. All training ran on Fireworks Serverless Training. Fireworks states it used its own data and no customer data.
Benchmark Results
Fireworks compared Ember-1 with Kimi K3 at three reasoning effort levels. Cost was computed with public Kimi K3 API pricing. These are Fireworks’ own published evaluations.
| Benchmark | N | K3 Low | K3 High | K3 Max | Ember-1 | Ember-1 vs K3 Max (cost) |
|---|---|---|---|---|---|---|
| Terminal Bench 2.1 | 89 | 76.4% | 77.6% | 80.9% | 82.0% | -51.9% / -23.1 USD |
| SWE-bench Verified | 500 | 80.4% | 86.0% | 93.2% | 92.2% | -15.5% / -68.1 USD |
| SWE-Interact | 75 | 6.7% | 13.3% | 21.3% | 20.0% | -32.5% / -60.8 USD |
| DeepSWE 1.1 | 113 | 55.8% | 62.8% | 66.4% | 75.2% | -23.7% / -126.9 USD |
| τ-2 Bench Airline | 50 | 64% | 64% | 64% | 66% | -5.9% / -0.3 USD |
Ember-1 leads K3 Max on Terminal Bench 2.1 and DeepSWE 1.1. It trails slightly on SWE-bench Verified and SWE-Interact. Across seven benchmarks and two customers’ production traffic, Fireworks says K3’s reasoning was shortened by 35 to 50% without sacrificing accuracy.
On Doximity’s Bedside Bench, a physician-validated set of 500 clinical cases, Ember-1 set a new cost-per-task Pareto frontier. That result comes from Fireworks’ new Specialized Intelligence Index.
Production A/B Test Results
Fireworks ran live A/B tests with 2 customers on production coding workloads. Both saw roughly 35% fewer tokens per task at comparable quality. In the published run, output tokens fell from 49.3K to 29.9K per task. Reasoning tokens dropped 71.3% and total tokens dropped 39%. The task score was essentially unchanged: 0.753 for Ember-1 versus 0.751 for K3. Average steps fell from 23.8 to 21.4. One customer now runs Ember-1 in production.
Ember-1 costs the same per token as Kimi K3 on Fireworks: $3.00 input, $0.30 cached input, and $15.00 output per 1M tokens. The savings come entirely from generating fewer tokens. At that output rate, the A/B figures work out to about $0.74 versus $0.45 in output cost per task (our calculation, output only).
Interactive Explainer
Credit: Source link
