Ember-1 cuts your agent's reasoning tokens ~40% at Kimi K3 quality

Fireworks retrained Kimi K3 to reason in fewer tokens — up to 71% fewer in A/B tests, cutting cost per task 39%. Live on OpenRouter now as a preview.

Nowline SEP 28 8:00 AM banner

Top AI stories from the last hour

Top AI stories from the last hour

Copy markdown

  • Same answers, ~40% fewer tokens

    Fireworks post-trained Moonshot's open-weight Kimi K3 to write shorter reasoning traces — roughly 40% fewer tokens across its evals (35–50% depending on the task) at comparable quality. Rates are unchanged at $3 / $15 per 1M in/out (K3's own prices), so every token it skips is money straight off your bill.

  • The agent-loop math

    In Fireworks' production A/B tests, output cost fell from $0.74 to $0.45 per task (39%) and average steps dropped 23.8 → 21.4 — meaning fewer tool calls and lower latency, not just cheaper tokens. That pencils out to about $8,700/month at 1,000 tasks/day. The savings compound in multi-turn loops; single-shot calls barely move.

  • Wins terminals, trails on SWE-bench

    It edges Kimi K3 on Terminal Bench 2.1 (82.0 vs 80.9), DeepSWE 1.1 (75.2 vs 66.4) and τ²-Bench Airline (66 vs 64), while sitting just under on SWE-bench Verified (92.2 vs 93.2). Treat it as a near-lateral quality move that pays for itself in tokens, not a strict upgrade.

  • How to try it before the window shuts

    It ships with a 1.05M-token context, MoE routing, function calling and image input; call it on OpenRouter (fireworks/ember-1), Fireworks' REST/Python, or any OpenAI-compatible client — a drop-in swap if you already run K3. It launched as a two-week serverless preview, reportedly made permanent on community demand, but with no permanence guarantee yet, so keep a fallback wired.