Fireworks Ember-1: Kimi K3 Quality With Fewer Tokens?

Fireworks AI released Ember-1 on September 23, 2026. It is the first model from Fireworks Research, and it is not a new base model. It is Moonshot's open-weight Kimi K3, post-trained on Fireworks' own training service to reason in fewer tokens. Fireworks says it delivers K3's quality with 35 to 50% shorter reasoning across seven benchmarks, and about 35% fewer tokens per task in live A/B tests on two customers' traffic.

The price is the point. Ember-1 costs exactly what Kimi K3 costs on Fireworks, $3 per million input tokens, $0.30 cached and $15 output. Nothing on the rate card changed. If the bill falls, it falls only because the model wrote less on the way to the answer.

Every number here is Fireworks' own. No independent replication exists at the time of writing, and the model is a research preview with a two-week serverless window unless Fireworks makes it permanent. Both facts shape what a team should do with it.

The interesting part is who made it. Fireworks is an inference provider. Its business is serving other people's models fast and cheaply. It has released models before, such as the open-weight Firefunction-v2 in June 2024, but Ember-1 is the first from Fireworks Research. Ember-1 is a provider using its position, which is that it sees real production traffic, to post-train an open model for the thing customers actually pay for: fewer tokens per completed task. Our post on the Kimi K3 release covers the base model this starts from.

Fireworks published five agentic benchmarks with Ember-1 against Kimi K3 at its maximum reasoning effort. The launch post also shows K3 at low and high effort; the table below keeps only the max column, since that is the comparison Fireworks makes in its opening claim. The cost column is Fireworks' own measurement of the change in cost to run each benchmark.

Read it honestly and it is a mixed table, not a sweep. Ember-1 scores higher on three rows. The DeepSWE gain of almost nine points is the only one that is clearly larger than sampling noise on these sample sizes. It scores lower on SWE-bench Verified, the one row with 500 tasks, and on SWE-Interact, which has 75. A one-point gap on 89 or 50 tasks is one or two tasks either way.

What is not mixed is the cost column. Every row costs less to run, from 6% on the airline benchmark to 52% on Terminal Bench. That is the claim Fireworks is actually making: roughly the same scores for noticeably less spend. The phrase in its opening sentence is K3's quality with 40% fewer tokens. We would not repeat "same quality" in our own voice on this evidence. We would say the scores move by amounts a team could live with while the token count moves by amounts it would notice.

Fireworks also published an A/B results table from its live tests on two customers' production coding workloads, without saying which customer it covers. The score barely moved and the token count fell by more than a third. The last row is our arithmetic at the published $15 output rate, and it is illustrative: it ignores input tokens, which are the same on both models at the same price, and it ignores cache hits.

Scale that up and the shape is clear. A team running 1,000 such tasks a day would spend about $740 a day on output tokens with K3 and about $450 with Ember-1, a difference near $8,700 a month, on the same rate card. The reasoning tokens fell by 71%, which suggests the model still writes a similar amount of visible answer and a much shorter private chain of thought. Fewer steps per task, 21.4 against 23.8, also means fewer tool calls and less wall-clock time.

The numbers are customer workloads as reported by the vendor, and Fireworks does not say which customer the table describes. Yours will differ, and the only way to find out is to run both models on the same task set and count the tokens.

This release is a clean demonstration of a rule we keep returning to. The price per million tokens on a rate card tells you almost nothing about what an agent costs, because agents differ in how many tokens they burn to finish a job. Two models at the same list price can differ by a third in cost per task. Two models at very different list prices can end up the same.

We saw the expensive version of this on September 22, when Grok 4.7 listed at under a third of Opus 5.5's output price and still cost about twice as much per task in Artificial Analysis' tests. We saw the cheap version in Muse Spark 1.3's clarifying questions, which Meta says cut tokens by about a quarter by asking before acting, although Artificial Analysis measured more input tokens per task. Ember-1 is the same lever pulled a third way: keep the price, keep the weights, train the model to stop thinking sooner.

For a team paying for long-reasoning agent runs, the practical metric is output tokens per completed task at an acceptable pass rate. Everything else on the rate card is an input to that number, not a substitute for it.

Fireworks describes its research releases as two-week serverless access, made permanent based on community demand. That is a fair deal for a preview and a poor foundation for production. The routing below assumes the route may disappear.