Fireworks Ember-1 API: a Kimi K3 reasoning model that thinks less

Ember-1 is a reasoning model that Fireworks' research team trained from Kimi K3 to think less without answering worse. It is on FastMetal as ember-1.

What Fireworks says

• About 40% fewer reasoning tokens than Kimi K3. In A/B tests in customer environments, Fireworks reports about 35% fewer tokens per task at similar quality. A reasoning model's cost per request follows its output tokens, so fewer reasoning tokens means a cheaper request.
• Built for coding and agents: 92.2% on SWE-bench Verified (Kimi K3 at its maximum setting: 93.2%) and 82.0% on Terminal Bench 2.1 (Kimi K3: 80.9%).

These are the vendor's figures.

What we confirmed

• Image input, tool calling and json_schema structured output, each in real calls through FastMetal.
• A 1,048,576-token context.
• It reasons even on short questions. Asked "日本の首都は？" ("What is the capital of Japan?"), it spent 43 reasoning tokens before answering "東京" ("Tokyo").

Calling it

Send ember-1 as the model to the same OpenAI-compatible endpoint as every other model:

curl https://api.fastmetal.ai/v1/chat/completions \
  -H "Authorization: Bearer $FASTMETAL_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "ember-1",
    "messages": [
      {"role": "user", "content": "Find and fix the bug in this function:\n\ndef mean(xs):\n    return sum(xs) / len(xs) - 1"}
    ]
  }'

How it is billed

Ember-1 is billed per million input and output tokens. Its per-token rates are higher than Kimi K3's, so it saves money only when the shorter reasoning more than makes up the difference. That is most likely on work where Kimi K3 reasons at length. Current rates are on the Ember-1 and Kimi K3 model pages, and the comparison page shows both models' answers to the same prompts.

Billing is prepaid: you load a balance by card, from $5, and requests stop when it runs out.

Research preview

Fireworks released Ember-1 as a research preview and said that continued availability depends on usage. If Fireworks withdraws it, it leaves FastMetal too, so keep a fallback model ready before you build production work on it.

Everything else we serve is in the model catalog, and the billing model is on the pricing page.