Fireworks Research launches Ember-1, a Kimi K3 derivative that cuts reasoning tokens by 40%

What Fireworks announced

Fireworks Research has released Ember-1, a specialized model built on Kimi K3 that the company says delivers the base model's quality while generating roughly 40% fewer tokens. The announcement, which surfaced on the front page of Hacker News, positions Ember-1 as the first in a planned series of specialized models from the Fireworks Research team, with future entries guided by developer demand. It is rolling out now as a serving option alongside the original Kimi K3.

The cost of thinking too much

According to Fireworks, reasoning models like Kimi K3 can spend more than 90% of their generated tokens on internal chain-of-thought rather than on the answer itself. That overhead is expensive in a single request, but it compounds in multi-turn agentic workloads: every turn replays earlier reasoning back to the model, so context grows roughly quadratically with the number of turns, and long traces from early turns get re-read and re-billed on every later call.

The obvious remedy — dialing down K3's reasoning-effort setting — did not work, the company says, because lower effort gave up too much quality. Fireworks' experiments instead concluded that K3's reasoning is far longer than tasks actually require, and that the surplus can be trimmed without changing the answer, provided the model is trained to reason efficiently rather than simply instructed to think less.

How it was built

Getting there took more than 50 training experiments and over 200 evaluations, plus new training algorithms aimed at shortening reasoning without losing accuracy. All of it ran on Fireworks' Serverless Training platform, which the company says let the team launch experiments without provisioning or managing GPUs. The training data spanned mathematics, coding, instruction following, conversation, search, tool use and software engineering, covering both standalone problems and extended interactions. Fireworks says the approach deliberately preserves useful self-reflection — revisiting an assumption or tracing an outcome back to an earlier decision — while cutting unproductive reasoning loops, and that the learned behavior also restrains token use on failed attempts.

Benchmarks and pricing math

Fireworks evaluated Ember-1 against its parent model, computing cost from K3's public API pricing: $3 per million uncached input tokens, $0.30 cached and $15 output. The reported industry results:

• Terminal Bench 2.1: Ember-1 scored 82.0% versus 80.9% for K3 at max effort, at 51.9% lower cost
• SWE-bench Verified: 92.2% versus 93.2%, 15.5% cheaper
• DeepSWE 1.1: 75.2% versus 66.4%, 23.7% cheaper — ahead on quality as well as cost
• SWE-Interact (7 samples): 20.0% versus 21.3%, 32.5% cheaper
• τ-2 Bench Airline: 66% versus 64%, 5.9% cheaper

Across every benchmark with more than 50 test samples, the company claims Ember-1 sits on or near the quality-versus-cost Pareto frontier, matching K3 at max effort at a fraction of the cost and strictly dominating the low-effort setting. Fireworks also compared the model against GPT-6 Astra, Claude Opus-5 and GLM 5.3 and says it led on the same frontier. On Doximity's Bedside Bench, a physician-validated set of 500 clinical cases that is part of the Specialized Intelligence Index Fireworks introduced the same week, the company reports a new cost-per-task frontier against models including GPT-5.6 Sol, GPT-6 Astra and Claude Opus 5.

Production traffic

Benchmarks aside, Fireworks ran live A/B tests with two customers on production coding workloads, reporting roughly 35% fewer tokens per task at comparable quality, with task completion, success scores and failure rates holding or improving. One customer has since moved Ember-1 into live production and plans to scale it until it fully replaces the base model. Internally, Fireworks swapped the model into its own developers' everyday coding tools and says nobody noticed the change, while reasoning tokens fell 71.3% and total tokens fell 39%. Across seven benchmarks and the two customers' traffic, the company says K3's reasoning could be shortened by 35–50% without sacrificing accuracy.

Why it matters

For teams running reasoning models at scale, token spend — not capability — is increasingly the binding constraint, especially in agentic pipelines where reasoning is replayed every turn. Ember-1 is a model-level answer to that cost: rather than prompting tricks or an effort dial, it bakes efficiency into the weights themselves, and the A/B results on real production traffic are the more credible part of the story. It also sketches a product category — specialized derivatives of strong base models, tuned for cost — that other inference providers can copy. Two caveats are worth holding onto: every number here comes from Fireworks' own announcement rather than independent evaluation, and the model doubles as a showcase for the company's Serverless Training business. If outside benchmarks confirm the claims, the practical takeaway for buyers is that the cheapest way to run Kimi K3 may be to run a derivative of it instead.