Ember-1: Fireworks' New Model Cuts Coding Tokens by 40%

Reasoning models can burn over 90% of their tokens just thinking before they answer. The actual answer is the tip of the cost iceberg. Fireworks Research just shipped Ember-1, a retrained version of Kimi K3 that cuts 35-50% of reasoning tokens while accuracy barely moves.

The problem: thinking models overthink

Reasoning models like Kimi K3 generate an internal chain of thought before producing a final answer. That's fine for a single request. It gets expensive fast in agentic workloads, where a model calls tools, reads results, and reasons across multiple turns — every prior turn's reasoning trace gets replayed (and re-billed) on every subsequent call. Context grows roughly quadratically with turn count.

Fireworks tried the obvious fix first: lower the reasoning effort setting. It didn't work — quality dropped too much. The real problem wasn't that the model thought too long, it's that it thought inefficiently. That's what pushed the team toward retraining instead of tuning a knob.

How Ember-1 was built

According to Fireworks' official blog post (published September 27, 2026), the Research team ran more than 50 training experiments and over 200 evaluations, all on Fireworks Serverless Training — no GPU provisioning, pay only for what runs.

The key insight: not all of K3's reasoning is wasted. Some of it is genuine self-reflection — revisiting an assumption, responding to feedback, tracing an outcome back to a decision — and that part actually helps the model recover from mistakes. Ember-1 was trained to preserve that behavior while cutting the excess, using training data spanning math, coding, instruction following, conversation, search, tool use, and software engineering.

The numbers: benchmarks and production

Fireworks published a comparison of Ember-1 against three K3 reasoning-effort tiers across five industry benchmarks:

Benchmark	K3 max	Ember-1	Cost delta
Terminal-Bench 2	80.9%	82.0%	-51.9% tokens, -$23.1
SWE-bench Verified	93.2%	92.2%	-15.5% tokens, -$68.1
SWE-Interact	21.3%	20.0%	-32.5% tokens, -$60.8
DeepSWE 1.1	66.4%	75.2%	-23.7% tokens, -$126.9
τ²-Bench Airline	64%	66%	-5.9% tokens, -$0.3

Worth noting: on DeepSWE 1.1, Ember-1 doesn't just cost less — it scores higher than K3 at max effort. That suggests trimming excess reasoning can sometimes sharpen focus instead of just saving money, by cutting off unproductive reasoning loops.

Fireworks also ran live A/B tests with two production customers on real coding workloads. Result: roughly 35% fewer tokens per task at comparable quality, with downstream metrics like task completion rate holding steady or improving. One customer has already moved Ember-1 into production, planning to fully replace the base model.

The internal validation is arguably the most telling data point. Fireworks quietly routed its own internal coding traffic to Ember-1 before any external announcement. Developers didn't notice the switch. For a model whose entire pitch is "same answers, fewer tokens," an invisible rollout is the strongest signal it works.

What this means for developers

If your team runs agentic coding workloads on Kimi K3 or similar reasoning models, costs tend to balloon non-linearly as tasks stretch across multiple turns — exactly because of the replayed-reasoning problem above. Ember-1 isn't magic, but it points at a real lever: instead of buying a bigger model or brute-forcing lower reasoning effort, you can retrain to strip wasted reasoning while keeping the part that actually generates value.

A few caveats worth weighing. Ember-1 is currently a "Research Preview" on Fireworks Serverless — free access for two weeks, after which Fireworks decides whether to make it permanent based on demand. It's not a fully independent model; it's still built on K3's underlying architecture, so K3's inherent limits (context window, multilingual capability, and so on) carry over. And every benchmark number here comes from Fireworks itself — no independent third-party verification exists yet at time of writing.

For teams actively optimizing AI agent costs, especially multi-turn workloads like automated code review or debugging loops, this is worth a two-week trial run before deciding whether to commit long-term.

What to know

• Ember-1 is a retrained Kimi K3, not an entirely new model
• Saves 35-50% of reasoning tokens while matching or beating K3 on most benchmarks
• Currently a free two-week Research Preview on Fireworks Serverless
• All benchmark numbers come from Fireworks itself — no independent verification yet
• Best fit for multi-turn agentic workloads, where replayed-reasoning costs compound fastest

---

Content assisted by AI (Amy 🌸). Reviewed by the author.