Fireworks Ember-1: Coding Model That Trims Reasoning Tokens
Fireworks AI published Ember-1 on its blog on 23 September 2026: not another frontier base model, but Moonshot Kimi K3 post-trained to stop paying for reasoning the task never needed — Fireworks says it generates roughly 40% fewer tokens at comparable quality, with production A/B work on one coding workload showing 71.3% fewer reasoning tokens
Tool Overview
Features, steps and FAQ below
Features
- ✓ Post-training approach: Fireworks says Ember-1 was post-trained by Fireworks Research from Moonshot Kimi K3, targeting coding and agentic workloads with the goal of K3-quality answers at K3 prices for fewer tokens
- ✓ Fewer reasoning tokens: Fireworks says it produces roughly 40% fewer generated tokens, with production A/B work on one coding workload showing 71.3% fewer reasoning tokens while quality scores held flat
- ✓ Specs: it supports image input and function calling with a 1M-token context window, priced at $3 per million input tokens, $15 output and $0.30 cache reads — the same rates as Kimi K3
- ✓ A specialised series: Fireworks says Ember is the start of a model series shaped for economics on real workloads rather than leaderboard maximum reasoning, and it argues that routing K3 together with Fable can beat either model alone on heterogeneous agent tasks
- ✓ Availability: offered as a research preview on Fireworks Serverless, suited to teams that previously had to pay for reasoning they did not need and want to run a small comparison first
How to Use
- Find the Ember-1 research preview on the Fireworks platform and call it with the same prompts as your existing workflow
- Run an A/B against Kimi K3, comparing reasoning-token consumption, latency and quality scores
- Assess the change in cost per completed task (not just per token) on your coding and agentic work
- Follow the Fireworks blog for the rest of the Ember series and its routing proposals
FAQ
What is Ember-1?
A research preview published by Fireworks AI on 23 September 2026, which Fireworks says is post-trained from Moonshot Kimi K3 for coding and agentic workflows with fewer unnecessary reasoning tokens.
How is it different from Kimi K3?
Fireworks says it is not a new base model but a post-trained K3 tuned to stop paying for reasoning the task never needed: roughly 40% fewer generated tokens at comparable quality, with 71.3% fewer reasoning tokens in one production coding workload.
What are the pricing and specs?
Pricing matches Kimi K3 at $3 per million input tokens, $15 output and $0.30 cache reads, with image input, function calling and a 1M-token context window.
What is it best for?
Fireworks positions it for coding and agentic workflows, particularly for teams whose long-horizon tasks were getting expensive because the model over-thought them.
What else does Fireworks propose?
Alongside Ember-1, Fireworks argues for routing across heterogeneous agent tasks, saying that combining K3 with Fable can beat either model on its own.