Ember-1: Teaching AI Agents to Reason Efficiently

Suppose you ask an AI agent—software that can plan, call tools, inspect results, and continue working—to fix a failing test. The final patch might be a few lines. Behind it, the model may spend thousands of tokens planning, revising, and narrating dead ends. Then the next turn sends much of that history back again.

That mismatch is the idea behind Ember-1, a specialized reasoning model from Fireworks Research. Released on September 23, 2026, Ember-1 is built on Kimi K3 and is reported to use about 40% fewer tokens while maintaining comparable quality across Fireworks evaluations. The interesting part is not that it thinks less by command; it was trained to decide which thinking earns its place. (fireworks.ai)

Tokens are the meter running in the background

A token is a small piece of text that a model processes. A short word may fit in one token, while a longer word, punctuation mark, or piece of code may be split into several. Reasoning models also generate intermediate planning tokens, whether or not an application shows that planning to the person using it.

An application programming interface, or API, usually reports input tokens and output tokens separately. Ember-1 is currently listed on Fireworks Serverless at $3 per million input tokens, $0.30 per million cached input tokens, and $15 per million output tokens. As an illustration, 10,000 output tokens cost about $0.15 at that rate; a 40% reduction to 6,000 tokens would cost about $0.09 before input and tool traffic are counted.

That difference sounds small until an agent performs thousands of coding tasks. A few cents saved per task can become a meaningful operating cost reduction, especially when the model is called repeatedly during a single job.

Why agents magnify wasted reasoning

An agentic workload is a loop: the model decides what to do, calls a tool, observes the result, and decides again. A coding agent might inspect files, edit a function, run tests, read the failure, revise the patch, and repeat several times.

If every turn sends the complete conversation back to the model, the amount of history being reread can grow roughly quadratically. Imagine each turn adding the same amount of text. The first call reads one unit, the second reads two, the third reads three, and so on. The total is proportional to 1 + 2 + 3 +... + n, not merely n.

A context window is the maximum amount of text a model can consider in one request. Ember-1 is currently listed with a context length of about 1.04 million tokens, but a large window does not make repeated reasoning free. Caching can reduce the price of repeated input, while long histories can still affect latency, memory use, and the room available for new information.

The training idea: keep reflection, cut the loops

Why not ask a reasoning model to think less? Fireworks says lower reasoning-effort settings on Kimi K3 reduced quality too much. The useful parts of a reasoning trace include checking an assumption, responding to feedback, and tracing an outcome back to an earlier decision. Waste looks more like repeating the same plan, exploring branches that no longer matter, or continuing after the answer is already supported.

Ember-1 was trained to preserve the first behavior while reducing the second. Fireworks describes more than 50 training experiments and over 200 evaluations across mathematics, coding, instruction following, conversation, search, tool use, and software engineering. The collection included both standalone problems and extended interactions. Task feedback guided on-policy planning, meaning the model practiced making decisions in the same situations where it would later need to act and recover from mistakes.

The distinction matters. Ember-1 is not merely Kimi K3 with a smaller output limit. A hard limit can cut off useful work. The goal here is learned token efficiency: the model develops a habit of reflecting when reflection helps, then stopping when more narration is unlikely to improve the result.

Measuring cost without pretending quality is one number

Fireworks evaluates Ember-1 through its Specialized Intelligence Index, or SII. The index focuses on domain benchmarks built around practical work and compares quality alongside cost and task duration. Its methodology also makes an important limitation clear: the index is not a measure of general intelligence, and its standardized harness may not represent the best possible setup for every model.

On Doximity’s Bedside Bench, a physician-validated evaluation covering 500 clinical cases across 10 specialized categories, Fireworks reports that Ember-1 reached a new cost-and-quality Pareto frontier. A Pareto frontier is the set of choices where no alternative is both cheaper and more accurate. That result is useful for model selection, but it is not permission to automate clinical decisions without appropriate safeguards.

The published industry results show the trade-off more clearly. Ember-1 scored 82.0% on Terminal Bench 2.1 compared with 80.9% for Kimi K3 at maximum reasoning, while scoring 92.2% on SWE-bench Verified compared with K3’s 93.2%. Fireworks also reports roughly 35% fewer tokens per task in two production coding A/B tests, where traffic was split between models and compared on real workloads. Internal developers reportedly continued their normal coding work without noticing the model change.

A small first experiment

Trying Ember-1 does not require redesigning an application. Fireworks provides a Python client with a chat-completions interface, and the model path is accounts/fireworks/models/ember-1.

from fireworks import Fireworks

client = Fireworks
response = client.chat.completions.create(
 model='accounts/fireworks/models/ember-1',
 messages=[
 {
 'role': 'user',
 'content': 'Summarize this bug report in five bullets.'
 }
 ],
)

print(response.choices[0].message.content)

For a fair comparison, keep the prompt, tools, sampling settings, and maximum number of turns the same. Measure more than output quality: record task completion, retries, tool-call count, wall-clock duration, and input and output tokens. An agent that uses fewer tokens but needs twice as many retries has not necessarily become cheaper. (docs.fireworks.ai)

Where Ember-1 fits

Ember-1 is not a universal replacement for a larger reasoning model. Unusual tasks, unfamiliar tool chains, and problems where one missed detail is expensive still deserve side-by-side testing against Kimi K3 or another strong model. Ember-1 is most compelling when the workload repeats, reasoning tokens dominate the budget, and success can be measured reliably.

The broader lesson reaches beyond one model release. Efficient AI is not only about faster hardware or lower-precision weights. A model that learns when to revisit an assumption—and when to stop revisiting it—can make long-running AI agents more affordable without turning off the intelligence that made them useful in the first place.