Overview
Ember-1 is a specialized kimi-k3 derivative from fireworks-ai Research — post-trained to emit shorter reasoning traces while maintaining benchmark parity. Research preview on Fireworks serverless.
Key Facts
- Released: September 23, 2026
- Base model: kimi-k3 (moonshot-ai)
- Token reduction: ~40% fewer total tokens; 35-50% shorter reasoning (vendor-reported)
- Pricing: Same as K3 — 15/M output
- Training: 50+ experiments, 200+ evaluations on Fireworks training stack
Benchmarks (vendor-reported)
| Benchmark | Ember-1 | Kimi K3 |
|---|---|---|
| SWE-bench Verified | 92.2% | 93.2% |
| Terminal Bench 2.1 | 82.0% | 80.9% |
Live A/B: ~35% token reduction at comparable quality; one enterprise customer in production.
All benchmarks vendor-reported; not independently reproduced. Research preview — permanence depends on demand.