AI Desk · AI Fingerprinting · Compare
Google: Gemma 2 27B · Google: Gemma 4 26B A4B
Identity similarity: lowDivergent fingerprints · 32/50These models tokenise differently. The rows that part here are the English prose, CJK, non-Latin script and emoji strings. Divergence on high-information strings separates unrelated vocabularies within a handful of samples.
Google: Gemma 2 27B recordGoogle: Gemma 4 26B A4B recordOpen in the interactive bench
Tokenizer fingerprint
32 of 50 mutually clean strings agree| Test string | Google: Gemma 2 27B | Google: Gemma 4 26B A4B | Verdict |
|---|---|---|---|
| en-prose | 14 | 14 | match |
| en-long | 15 | 16 | differs |
| spaces-20 | 3 | 3 | match |
| spaces-60 | 4 | 4 | match |
| tabs-20 | 3 | 3 | match |
| newlines-20 | 3 | 3 | match |
| mixed-ws | 9 | 9 | match |
| digits-9 | 9 | 9 | match |
| digits-12 | 12 | 12 | match |
| digits-30 | 30 | 30 | match |
| digits-sep | 12 | 12 | match |
| float-long | 22 | 22 | match |
| zh-common | 10 | 10 | match |
| zh-long | 13 | 12 | differs |
| zh-rare | 21 | 25 | differs |
| ja-kana | 7 | 7 | match |
| ja-kanji | 9 | 9 | match |
| ko | 12 | 11 | differs |
| ru | 12 | 12 | match |
| ar | 11 | 10 | differs |
| he | 12 | 13 | differs |
| hi | 13 | 8 | differs |
| th | 12 | 10 | differs |
| el | 14 | 13 | differs |
| emoji-basic | 5 | 5 | match |
| emoji-skin | 9 | 10 | differs |
| emoji-zwj-family | 7 | 7 | match |
| emoji-zwj-x3 | 21 | 21 | match |
| emoji-flags | 7 | 8 | differs |
| emoji-prof | 11 | 11 | match |
| math | 26 | 26 | match |
| boxdraw | 17 | 17 | match |
| combining | 9 | 10 | differs |
| cjk-ext-b | 16 | 16 | match |
| surrogates | 18 | 27 | differs |
| zalgo | 22 | 22 | match |
| rtl-mix | 6 | 5 | differs |
| py-code | 28 | 28 | match |
| py-indent | 18 | 18 | match |
| json | 28 | 28 | match |
| html | 17 | 17 | match |
| regex | 46 | 44 | differs |
| camel | 6 | 6 | match |
| snake | 11 | 11 | match |
| rare-word-x5 | 25 | 25 | match |
| repeat-tok | 20 | 21 | differs |
| base64 | 33 | 34 | differs |
| hex | 15 | 16 | differs |
| url | 23 | 23 | match |
| uuid | 36 | 36 | match |
Values are the marginal prompt-token cost of each string with the model's fixed overhead subtracted. An x marks a row excluded as corrupt for that model.
API surface
The declared serving contracts differ| Field | Google: Gemma 2 27B | Google: Gemma 4 26B A4B | Verdict |
|---|---|---|---|
| Context length | 8192 | 262144 | × differs |
| Max output tokens | 2048 | – | × differs |
| Supported parameters | frequency_penalty, max_tokens, presence_penalty, repetition_penalty, response_format, seed, stop, structured_outputs, temperature, top_p | frequency_penalty, include_reasoning, logit_bias, logprobs, max_tokens, min_p, presence_penalty, reasoning, repetition_penalty, response_format, seed, stop, structured_outputs, temperature, tool_choice, tools, top_k, top_logprobs, top_p | × differs |
| Defaults | {} | {"temperature":1,"top_p":0.95,"top_k":64} | × differs |
| Reasoning contract | – | {"mandatory":false,"default_enabled":false} | × differs |
| Modalities | text->text | text+image+video->text | × differs |
Google: Gemma 4 26B A4B : OpenRouter's listing shows 235,929 output tokens, which is 90% of the context window, the figure it carries where the provider declares no maximum.
Method: tokenizer fingerprinting, v1, 50 strings. This page is static and dated; the interactive bench compares any two of the 338 measured models.