Ember-1 and LLM Cost Optimization: Fireworks' Own Table Adds Up to 23%, Not Half

Fireworks published Ember-1 on 23 September: Kimi K3, retrained so that it reasons for fewer tokens and reaches the same answers. For anyone paying for agentic coding by the token, this is LLM cost optimization in the literal sense: less work, smaller invoice. The question is by how much, and the announcement gives several answers.

What shipped

The Ember-1 model page lists it as a mixture-of-experts model built on K3, with 2.78T parameters, a 1040k-token context, function calling and image input, created on 22 September. Its serverless price is $3.00 per million input tokens, $0.30 cached, and $15.00 output. The K3 page on the same site lists exactly the same three numbers. With identical rates, any cost reduction comes from token counts alone.

Fireworks calls it a Research Preview. The post describes research releases as two weeks of serverless access, with a model made permanent if the community asks for it.

How big the saving is depends on which sentence you read

The model page says approximately 40% fewer tokens. The post's first heading says half. Its body says reasoning can be shortened by 35 to 50% across seven benchmarks and two customers. The live A/B tests with two customers came in at about 35% fewer tokens per task. The A/B table shows output tokens falling from 49.3K to 29.9K, which the post reports as a 39% total cut (71.3% of reasoning tokens alone).

The most useful number is one the post never prints. Fireworks costed five benchmarks against K3 at its default max effort using the rates above, and gave each row a percentage and a dollar saving:

• Terminal Bench 2.1, 89 tasks: 51.9%, $23.10
• SWE-bench Verified, 500 tasks: 15.5%, $68.10
• SWE-Interact, 75 tasks: 32.5%, $60.80
• DeepSWE 1.1, 113 tasks: 23.7%, $126.90
• τ-2 Bench Airline, 50 tasks: 5.9%, $0.30

Divide each dollar figure by its percentage to recover the K3 baseline, then sum. K3 at max costs about $1,211 across the five runs, and Ember-1 saves $279 of it, so the saving is about 23%. That arithmetic assumes each row's percentage and dollar figure describe the same cost change, which is how the column reads. The half-cost result comes from Terminal Bench, the cheapest row, at about $44 of baseline. DeepSWE and SWE-bench Verified carry more than 80% of the spend, and their cuts are 23.7% and 15.5%.

Quality mostly holds. Ember-1 scores higher on three rows, including DeepSWE at 75.2% against 66.4%. It scores lower on two: SWE-bench Verified, 92.2% against 93.2%, and SWE-Interact, 20.0% against 21.3%. The post's Pareto claim covers only benchmarks with more than 50 samples, and τ-2 Airline has exactly 50. That row also has the smallest cut.

The baseline is Fireworks' K3, not Moonshot's

Moonshot's own K3 model card reports 88.3 on Terminal-Bench 2.1 and 67.5 on DeepSWE, both run with its Kimi Code harness. Fireworks' K3-at-max row shows 80.9 and 66.4. The Ember-1 post does not name the harness behind its table. The Specialized Intelligence Index methodology, which the post points to, says the index runs one pinned open-source harness per benchmark. It also says the index does not measure the best a model can do, since a purpose-built harness usually scores higher. So every relative number above compares two models inside Fireworks' setup. Your harness is a third setup, and nobody has measured it yet.

The K3 card adds one more constraint. K3 was trained to have its full previous reasoning passed back on every turn, and it always thinks, with an effort setting of low, high or max (max by default). The Ember-1 post argues that this replayed reasoning is what makes long agent sessions expensive. By that same argument, the saving should grow with session length, so a short-session workload will see less of it than a long agent loop.

What it breaks

Moving from K3 to Ember-1 is a model-ID change on Fireworks and nothing more. The limits show up in what the Ember-1 page leaves out, compared with the K3 page next to it:

• No weights. K3's page links to moonshotai/Kimi-K3 on Hugging Face, whose license permits derivative works. Ember-1's page links to no repository, and the newest model repository in the fireworks-ai organization on Hugging Face dates from October 2024. If you self-host K3, Ember-1 is not available to you at all.
• No fine-tuning. The Ember-1 page lists fine-tuning as not supported, while K3's supports LoRA. The announcement says training support for Ember-1 is launching, so check the model page, not the post.
• Serverless only. K3's page offers on-demand deployment and Ember-1's does not. Fireworks' lifecycle docs promise at least two weeks' notice before a serverless model is removed, and recommend on-demand deployments for production workloads that need a stable model. For Ember-1, that option does not exist yet.

Verdict: test it, do not migrate to it

A 23% cut at equal price on long agentic coding is a real saving, and roughly a quarter is a fairer planning number than half. But it is a two-week preview that you cannot pin, fine-tune or take in-house. That makes it an experiment to run, not a platform decision. As we argue in From PoC to Production, routing to a cheaper model counts as cost control only when an eval you can defend says quality held. Fireworks measured its benchmarks, not your traffic.

The claim this post stakes: on production agentic coding traffic, the dollar-weighted saving will land nearer 25% than 50%. An independent replay of SWE-bench Verified or DeepSWE showing a cost cut of 40% or more at equal pass rate would prove it wrong.

Run this before the preview window closes:

1. Replay at least 100 of your own multi-turn tasks, not 50, through both accounts/fireworks/models/kimi-k3 and accounts/fireworks/models/ember-1. Pass back the full assistant messages on every turn, as the K3 card requires.
2. Take the prompt_tokens and completion_tokens Fireworks returns on each response, price them at the posted rates, and sum the dollars per model. Do not average per-task percentages, which lets a cheap task count as much as an expensive one.
3. Switch only if the dollar saving clears 20% and your pass rate stays within one point of K3. Until Ember-1 offers on-demand deployment, keep K3 as a one-line fallback.

Deciding which model a workload should run on, and what that costs, is the core of an LLM consulting engagement. The measurement above is the first step.