Ember-1

Ember-1 is a language model from Fireworks AI, published on the company’s blog on 23 September 2026 and listed independently by OpenRouter and Vercel’s AI Gateway within hours. It is not a new base model: Fireworks retrained Moonshot AI’s Kimi K3, a 2.78-trillion-parameter mixture-of-experts model already on this board, to reach the same answers while writing less reasoning. The developer’s own tests put the saving at about 40 per cent fewer reasoning tokens, and its table shows mixed results against Kimi K3: slightly ahead on Terminal Bench 2.1 (82.0 against 80.9 per cent) and DeepSWE 1.1, slightly behind on SWE-bench Verified (92.2 against 93.2). Those are Fireworks’ own measurements and are not stored as scores; no independent evaluation had been published when this was read on 25 September 2026. It takes text and images and returns text, with function calling and structured output. Context length is given as about a million tokens by all three listings, though they differ slightly (1,040K on Fireworks’ page, 1,048,576 on OpenRouter and Vercel). The price is $3.00 per million input tokens, $0.30 cached and $15.00 per million output tokens — the same rates Fireworks charges for Kimi K3, so the saving shows up as fewer billed output tokens rather than a lower rate. Two things matter for a buyer: Fireworks calls it a research preview with two-week serverless access, so it may be withdrawn or repriced, and the weights are not published.