Fireworks Research published Ember-1 September 23, 2026 — specialized Kimi K3 derivative holding accuracy while spending ~40% fewer tokens. Research preview on Fireworks serverless; two-week access window.
50+ training experiments, 200+ evaluations. SWE-bench Verified: 92.2% (vs K3 93.2%). Terminal Bench 2.1: 82.0% (vs K3 80.9%). Live A/B: ~35% token reduction at comparable quality; one customer in production.
All figures vendor-reported; not independently reproduced. Same pricing as Kimi K3: 15/M output.