Ember-1: Fireworks post-trains Kimi K3 to cut tokens ~40%

Same $3/$15 rate card and 1M context as K3, but ~71% less reasoning per task. Live now on Fireworks, OpenRouter and Vercel's gateway.

Nowline SEP 28 12:00 PM banner

Top AI stories from the last hour

Top AI stories from the last hour

Copy markdown

  • Kimi K3 quality, minus the padding

    Ember-1 is Fireworks' post-trained Kimi K3 that holds the same quality on ~40% fewer tokens (35–50% across seven benchmarks). It even edges K3 on Terminal Bench 2.1 (82.0% vs 80.9%) and DeepSWE 1.1 (75.2% vs 66.4%), trailing only a hair on SWE-bench Verified (92.2% vs 93.2%).

  • The number that matters is tokens-per-task

    In a production A/B test, reasoning tokens fell 71% and output dropped from 49.3K to 29.9K per task at matched quality. Pricing is unchanged from K3 ($3 in / $15 out per 1M), so a shop running ~1,000 tasks a day could shave roughly $8,700/month off output alone — measure the tokens you burn, not the rate card.

  • It's a drop-in today, three ways

    Call `fireworks/ember-1` on Fireworks Serverless, OpenRouter, or Vercel's AI Gateway (`vercel ai-gateway setup`), through the OpenAI Chat/Responses or Anthropic Messages APIs. You get a 1M-token context, text + image input, tool calling, implicit caching, and zero data retention with no training on the endpoint.

  • The catch: don't wire it into prod yet

    It's an API-only research preview on a two-week serverless window, with no weights released and no self-hosting. Shadow-test it against K3 on your real traffic before you commit — the route may not outlast the preview.