Why high: an exact fingerprint shared with 5 models that declare the same family, without a second signal of a different kind yet.
Qwen: Qwen3.8 27B is Alibaba (Qwen)'s model. Its measured tokenizer sits with the Qwen group, which is evidence that it builds on the Qwen tokenizer. Tokenizer reuse is normal engineering, open vocabularies travel between labs, and this says nothing against Alibaba (Qwen)'s authorship of the model itself.
Re-measured on 5 October 2026 on four providers one at a time: Reka, CoreWeave, Parasail and Chutes. The 21 August reading was taken while OpenRouter spread the calls across eight providers, and five of its rows (English prose, Hindi, the emoji sequence, the mathematical-alphabet string and the UUID) came back four or five tokens below what every provider reports on its own. The four single-provider readings agree with each other, and the corrected record matches Qwen 3.6 27B and Qwen 3.6 35B exactly on all 50 strings: Qwen 3.8 27B ships the Qwen 3.6 vocabulary. The record counts that as a changed fingerprint because the August numbers were published; the vocabulary behind the name did not move.
| We checked | What we found | Where it came from |
|---|---|---|
| How it counts tokens | Counts every one of the 50 test strings exactly like 5 other models | tk_06ffa14a |
| How its API is set up | A combination of settings no other listing in the catalogue uses | api_e620ed95 |
| How much it can read and write | Reads up to 1,000,000 tokens at once, writes up to 131,072 | its own listing |
| Whether it thinks before answering | Reasons when asked to, at xhigh, medium, low, xhigh by default | its own listing |
| Who served our requests | Reka | we measured it, 5 October 2026 |
| Measured twice | Between 21 August 2026 and 5 October 2026, 5 of 50 strings changed count in a pattern no template explains: a different tokenizer now answers under this name. | 45/50 agree |
Click a row to open the full comparison.
We sent Qwen: Qwen3.8 27B 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: 52 prompt tokens.
Qwen3.8 27B is an open-weight dense vision-language model from Qwen. It is suited for coding, professional workflows, research, multimodal interaction, and long-running agent tasks, with flexible thinking that can be...
Supported parameters
frequency_penaltyinclude_reasoninglogit_biaslogprobsmax_tokensmin_ppresence_penaltyreasoningreasoning_effortrepetition_penaltyresponse_formatseedstopstructured_outputstemperaturetool_choicetoolstop_atop_ktop_logprobstop_p
defaults: {"temperature":1,"top_p":0.95,"top_k":20}
