Ember-1 Guide: Kimi K3 Quality With Fewer Reasoning Tokens

The usual way to make a reasoning model cheaper is to turn its effort setting down. You save tokens, and you lose accuracy on the hard tasks. Ember-1 is a bet that this trade is not forced: that a large share of a reasoning model's thinking is redundant, and that training can remove the redundant part while keeping the part that matters.

Fireworks Research published Ember-1 on 23 September 2026. It is built on Moonshot AI's Kimi K3, and the headline claim is that it delivers K3's quality with roughly 40% fewer tokens. Everything below is about how much of that claim you should believe before running it yourself.

What Fireworks actually changed

Ember-1 is not a new pre-trained model. It is Kimi K3 with further training aimed at one behaviour: reasoning length. Fireworks says it trained on a broad task mix — maths, coding, instruction following, conversation, search, tool use and software engineering — with task feedback driving on-policy learning, and with an explicit goal of keeping self-reflection while cutting unproductive reasoning. It says it used its own data and no customer data.

That means the underlying shape is K3's. Kimi K3 is a 2.8T-parameter mixture-of-experts model with about 104B active per token and a 1M-token context window. If you need the background on what that buys and costs, the Kimi K3 guide covers it. Ember-1 inherits the window and, per the catalogue listings, accepts text and image input with tool calling and structured outputs.

What Fireworks has not published is equally relevant: there is no statement about weights or a licence for Ember-1 itself. Treat it as a hosted model, not an open-weight one, until that changes. K3's own weights ship under Moonshot's custom commercial terms, so a derivative would not be permissive by default anyway.

The benchmark table, read carefully

Fireworks compares Ember-1 against K3 at three effort levels — low, high and max, where max is K3's default. On its own table:

• Terminal-Bench 2.1 (89 tasks): Ember-1 82.0% against K3 Max 80.9%.
• SWE-bench Verified (500 tasks): 92.2% against 93.2%.
• DeepSWE 1.1 (113 tasks): 75.2% against 66.4%.
• SWE-Interact (75 tasks): 20.0% against 21.3%.
• τ-2 Bench Airline (50 tasks): 66% against 64%.

Two things stand out. First, Ember-1 is not ahead everywhere; it trails K3 Max slightly on SWE-bench Verified and SWE-Interact. That is a more credible result than a clean sweep. Second, several of these sets are small. A one-point gap on 89 tasks is a single task. The meaningful claim is not that Ember-1 is smarter than K3, but that it lands at roughly the same accuracy while spending less. Fireworks prints a reduction against K3 Max for each benchmark, from about 6% on τ-2 Airline to about 52% on Terminal-Bench, alongside a dollar figure computed from K3's public API prices, though the table does not label whether the percentage is tokens or cost.

All of these are vendor-reported and none had been independently reproduced when it launched. Vendor-claimed vs independent benchmarks covers how to weight numbers like these.

Where the saving shows up

The most useful evidence in the announcement is not the benchmark table but a customer A/B test on a production coding workload. Fireworks reports the score moving from 0.751 to 0.753, average steps falling from 23.8 to 21.4, and output tokens per task dropping from 49.3K to 29.9K — a 39% reduction in total tokens, driven by a 71% cut in reasoning tokens. Two customers, it says, saw about 35% fewer tokens per task at comparable quality.

For an agent, fewer reasoning tokens pays twice. You pay less for output on the turn itself, and you carry less into later turns if your harness keeps reasoning in the transcript. Fewer steps per task also means fewer tool calls and less wall-clock time. Hidden reasoning output is usually the line item that surprises people.

Note what does not change: the price per token. OpenRouter lists Ember-1 at $3 input and $15 output per million tokens at the time of writing, with cached input at $0.30 — the same rates Fireworks quotes for Kimi K3's public API. The saving is entirely in volume, so it only exists if the token reduction holds on your workload.

Research preview means plan for change

Fireworks ships Ember-1 as a research preview, a serving option alongside base K3 on its serverless platform. The same announcement introduced two-week serverless access for research models, which become permanent based on community demand. OpenRouter lists it as fireworks/ember-1.

That has practical consequences. Do not hard-wire a preview model into a pipeline with no fallback. Route to it with K3 behind it, so a withdrawn endpoint degrades cost rather than availability. Model routing and fallbacks covers the pattern. If you do adopt it, log the exact model identifier your provider reports on each response, so you notice if what serves the name changes while you are measuring it.

How to test it against K3

The claim is specific enough to test cheaply, because the comparison is the same base model with and without the extra training.

1. Pick fifteen to twenty real tasks from your own history — the kind your agent actually runs, not puzzles.
2. Run each through K3 at its default effort and through Ember-1, on the same harness and starting state.
3. Record pass or fail, steps, output tokens and reasoning tokens separately where your provider reports them.
4. Also run K3 at a lower effort. If K3 low matches Ember-1 on your tasks, the extra training is not buying you anything.
5. Compare cost per completed task, not cost per token.

How to benchmark LLMs on your own work has the harness details.

Decision rule

• Try Ember-1 if you already run Kimi K3 on long agent loops and output tokens dominate the bill.
• Stay on K3 if you need a model that will still be there next quarter with certainty, or if your tasks are short enough that reasoning length is not the cost driver.
• Look elsewhere if the real problem is K3's price tier itself. A 40% cut on a $15 output rate is still more expensive than a cheaper model that solves the task.