Fireworks took Kimi K3 and trained it to stop wasting reasoning tokens: 40% fewer tokens, same answers, available as a two-week Research Preview on serverless. The interesting part is they didn't just dial down reasoning effort — K3's effort knobs gave up too much quality — they trained it across math, coding, tool use, and agentic tasks, with feedback teaching it which reasoning to keep. 50+ training experiments, 200+ evaluations. The mechanism matters for agents: reasoning models spend over 90% of output tokens on thinking, and in multi-turn agentic workloads prior traces get replayed and re-billed each turn, so context grows roughly quadratically. Cutting the trace cuts the quadratic's coefficient. Numbers are mostly solid: Terminal Bench 2.1 at 82% beats K3-max's 80.9% at 51.9% less cost; SWE-Bench Verified 92.2% vs 93.2% (-15.5% cost); the weak spot is SWE-Interact, 20.0% vs 21.3%. Live A/B with two customers: ~35% fewer tokens at comparable quality, one customer now in full production. Their internal validation is "developers didn't notice," which is either the strongest signal or a commentary on how closely developers read agent traces — the source doesn't say which. The two-week serverless preview means this could vanish by the time you've integrated it. Plan your benchmark run accordingly.