Fireworks Ember-1: Kimi K3 Quality on 40% Fewer Tokens

Ember-1 is Fireworks' post-trained version of Moonshot's open-weight Kimi K3, built to land on nearly the same answer while reasoning through far fewer tokens to get there. Announced on Fireworks' blog on September 23, 2026, it holds within about a point of Kimi K3's own scores while cutting token use by 15.5% on SWE-bench Verified and 51.9% on Terminal Bench 2.1.

• Ember-1 is built on Kimi K3, Moonshot AI's open-weight model, rolled out as a Research Preview on Fireworks Serverless on September 23, 2026.
• Fireworks reports 92.2% on SWE-bench Verified, 82.0% on Terminal Bench 2.1, and 75.2% on DeepSWE 1.1, using 15.5% to 51.9% fewer tokens than Kimi K3 Max, all on its own harness.
• Pricing matches the public Kimi K3 rate: $3 per million input tokens, $15 per million output, so fewer tokens per task means a smaller bill.
• Our AI Coding Leaderboard scores Kimi K3 at 93.4% on SWE-bench Verified via vals.ai, close to Fireworks' own 93.2% baseline for K3 Max, an unusually tight agreement between vendor and independent numbers.

What did Fireworks actually change?

Ember-1 is not a new foundation model. It's Kimi K3 run through extra reinforcement learning aimed at one behavior: cutting reasoning tokens that don't change the final answer. Fireworks says the work came from 50-plus training experiments and 200-plus evaluations across math, coding, tool use and software engineering. It also claims a new Pareto frontier on Doximity's Bedside Bench, a non-coding benchmark outside our lane, so treat that one as a training-method data point, not a coding claim.

RelatedCognition's SWE-2 Nearly Matches GPT-6 Astra for a Quarter the Price

How much does it actually save?

The cuts scale with how much reasoning a task needs. SWE-bench Verified barely moves: 92.2% against Kimi K3 Max's 93.2%, for 15.5% fewer tokens. Terminal Bench 2.1, where an agent runs long shell sessions, sees the biggest cut: 82.0% against 80.9% for 51.9% fewer tokens. DeepSWE 1.1 lands at 75.2% against 66.4%, a gain, for 23.7% fewer tokens. In live A/B production traffic, Fireworks reports roughly 35% fewer tokens per task at comparable quality, the number that matters most to anyone routing real traffic.

The catch is availability. Ember-1 launched as a Research Preview with a two-week window of guaranteed serverless access. Fireworks says continued availability past that depends on how much developers actually use it.

1. Jul 27, 2026Kimi K3 ships open weights. Moonshot AI releases 1.56TB of weights in 4-bit MXFP4.
2. Sep 23, 2026Fireworks launches Ember-1. Research Preview, post-trained from Kimi K3.
3. ~Oct 7, 2026Two-week preview window closes. Availability depends on usage, per Fireworks.

Does the math hold up against our own numbers?

This one is more honest than most vendor-versus-independent matchups. Fireworks' own baseline for Kimi K3 Max on SWE-bench Verified is 93.2%, and our independently verified figure, via vals.ai's bash-only mini-swe-agent harness, is 93.4%. Two unrelated pipelines landing 0.2 points apart makes the shared K3 baseline credible, and puts the roughly 1.2-point gap between K3 Max and Ember-1 close to real. It doesn't make Ember-1's 92.2% directly comparable to our 93.4%, since Ember-1 has only run on Fireworks' harness, and it doesn't extend to Terminal Bench: our 88.3% for Kimi K3 is a different test than Fireworks' Terminal Bench 2.1, where K3 Max scores 80.9%. Vendor scores on our board have historically landed 2.6 to 11.6 points above neutral harnesses, so read this gap as a good sign, not a verified one.

What does this actually cost, in practice?

Ember-1 bills at the same public rate as Kimi K3: $3 per million input tokens, $15 per million output. Price per token is unchanged, so savings come entirely from using fewer tokens per job. Assuming output tokens scale with reasoning length, a 15.5% cut means roughly a 15.5% cheaper SWE-bench-style fix, and a 51.9% cut roughly halves the cost of a Terminal Bench-style agent task. That's an assumption, not a guarantee: real workloads mix short and long reasoning turns. Run your own numbers with our AI cost calculator before committing budget.

RelatedKimi K3 Is Third Best at Coding. It Is Also Not Open.

What does this mean for the market?

Fireworks is a private inference company, not a foundation-model lab, and it just showed that big cost cuts don't require a new base model. A focused post-training pass on someone else's open weights got most of the way there. Moonshot benefits too: every dollar Fireworks saves running K3 is a reason to keep it in rotation over a closed API. Closed frontier labs should read this as pressure: a third party just cut inference cost this much without retraining anything, so the cost floor for coding-agent work keeps dropping regardless of who owns the weights. None of this is investment guidance, just a read on where the margin is heading.

Our take

The frontier that matters for coding models right now isn't the top-line score. Most serious contenders already cluster within a few points of each other on SWE-bench Verified, and the gap that shows up on an invoice is how many tokens it took to get there. Ember-1 is Fireworks betting that gap can be closed with post-training instead of a new model, and the numbers we can check hold up better than most vendor claims we've reviewed. What we'd flag is durability: Ember-1 lives entirely on Fireworks' infrastructure, tied to a two-week preview and to Moonshot shipping open weights worth post-training. If both hold, this looks like a real second axis of competition alongside raw accuracy. If either slips, it was a well-timed demo, not a lasting product.

Original analysis by GenZTech. Source: Fireworks AI.