AI Desk · AI FingerprintingNew arrival
Why high: an exact fingerprint shared with 4 models.
Fireworks: Ember-1 is Fireworks's model. Its measured tokenizer sits with the Moonshot AI group, which is evidence that it builds on the Moonshot AI tokenizer. Tokenizer reuse is normal engineering, open vocabularies travel between labs, and this says nothing against Fireworks's authorship of the model itself.
Listed on OpenRouter at 00:07 UTC on 24 September 2026 and measured the same day. Fireworks describes Ember-1 as a reasoning model from Fireworks Research built on Moonshot AI's Kimi K3, trained to reach Kimi K3's quality with about 40% fewer tokens. Its token counts are Kimi K3's on all 50 strings, measured with both models served by Fireworks, and the two also agree on all 18 special-token probes and carry the same 88-token chat template. Kimi K2.6, K2.7 Code and K2 Thinking match all 50 as well, because Moonshot has kept one vocabulary from K2 to K3; the median across 288 other measured models is 14 of 50. Two passes agreed on every string and the fixed overhead held at 88 tokens before and after. A model trained on top of another keeps its base's vocabulary, so the counts confirm the Kimi K3 base Fireworks names; the reasoning savings are Fireworks' own evaluations. Raw measurements and the working are in kept with the desk's research files.
Fireworks: introducing Ember-1 (23 September 2026)Ember-1 on OpenRouter
| We checked | What we found | Where it came from |
|---|---|---|
| How it counts tokens | Counts every one of the 50 test strings exactly like 4 other models | tk_ea6068bf |
| How its API is set up | A combination of settings no other listing in the catalogue uses | api_d77aa67f |
| How much it can read and write | Reads up to 1,048,576 tokens at onceOpenRouter's listing shows 943,718 output tokens, which is 90% of the context window, the figure it carries where the provider declares no maximum. | its own listing |
| Whether it thinks before answering | Reasons when asked to, at max, high, low, max by default | its own listing |
| Who served our requests | Fireworks | we measured it, 24 September 2026 |
Click a row to open the full comparison.
We sent Fireworks: Ember-1 fifty short pieces of text and recorded what each one cost it in tokens. The bars below are those costs. Two models built on the same tokenizer produce the same bars; a model built on a different one produces a different set, which is what makes this a fingerprint.
Overhead subtracted: 88 prompt tokens.
Ember-1 is a specialized reasoning model from Fireworks Research, built on [Kimi K3](https://openrouter.ai/moonshotai/kimi-k3). It is designed to make every token go further: it produces shorter reasoning traces, using roughly 40%...
Supported parameters
frequency_penaltyinclude_reasoninglogit_biaslogprobsmax_tokenspresence_penaltyreasoningreasoning_effortrepetition_penaltyresponse_formatstopstructured_outputstemperaturetool_choicetoolstop_ktop_logprobstop_p
defaults: {}
