AI Desk · AI Fingerprinting · Compare
DeepSeek: DeepSeek V3.2 Exp · DeepSeek: DeepSeek V4 Flash Vision Exp
Identity similarity: highExact tokenizer agreement · 50/50Exact agreement across every measured string is one strong signal of shared tokenizer lineage. The published settings differ, so the second signal the methodology needs for very high is absent here; the table below shows where they part.
DeepSeek: DeepSeek V3.2 Exp recordDeepSeek: DeepSeek V4 Flash Vision Exp recordOpen in the interactive bench
Tokenizer fingerprint
50 of 50 mutually clean strings agree| Test string | DeepSeek: DeepSeek V3.2 Exp | DeepSeek: DeepSeek V4 Flash Vision Exp | Verdict |
|---|---|---|---|
| en-prose | 14 | 14 | match |
| en-long | 16 | 16 | match |
| spaces-20 | 3 | 3 | match |
| spaces-60 | 3 | 3 | match |
| tabs-20 | 4 | 4 | match |
| newlines-20 | 4 | 4 | match |
| mixed-ws | 7 | 7 | match |
| digits-9 | 3 | 3 | match |
| digits-12 | 4 | 4 | match |
| digits-30 | 10 | 10 | match |
| digits-sep | 7 | 7 | match |
| float-long | 9 | 9 | match |
| zh-common | 9 | 9 | match |
| zh-long | 10 | 10 | match |
| zh-rare | 18 | 18 | match |
| ja-kana | 9 | 9 | match |
| ja-kanji | 11 | 11 | match |
| ko | 13 | 13 | match |
| ru | 13 | 13 | match |
| ar | 10 | 10 | match |
| he | 14 | 14 | match |
| hi | 18 | 18 | match |
| th | 11 | 11 | match |
| el | 16 | 16 | match |
| emoji-basic | 10 | 10 | match |
| emoji-skin | 14 | 14 | match |
| emoji-zwj-family | 11 | 11 | match |
| emoji-zwj-x3 | 33 | 33 | match |
| emoji-flags | 16 | 16 | match |
| emoji-prof | 19 | 19 | match |
| math | 22 | 22 | match |
| boxdraw | 27 | 27 | match |
| combining | 10 | 10 | match |
| cjk-ext-b | 16 | 16 | match |
| surrogates | 33 | 33 | match |
| zalgo | 36 | 36 | match |
| rtl-mix | 6 | 6 | match |
| py-code | 25 | 25 | match |
| py-indent | 15 | 15 | match |
| json | 20 | 20 | match |
| html | 16 | 16 | match |
| regex | 47 | 47 | match |
| camel | 7 | 7 | match |
| snake | 8 | 8 | match |
| rare-word-x5 | 26 | 26 | match |
| repeat-tok | 41 | 41 | match |
| base64 | 34 | 34 | match |
| hex | 12 | 12 | match |
| url | 17 | 17 | match |
| uuid | 27 | 27 | 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 | DeepSeek: DeepSeek V3.2 Exp | DeepSeek: DeepSeek V4 Flash Vision Exp | Verdict |
|---|---|---|---|
| Context length | 163840 | 1048576 | × differs |
| Max output tokens | – | 262144 | × differs |
| Supported parameters | frequency_penalty, include_reasoning, max_tokens, reasoning, response_format, structured_outputs, temperature, tool_choice, tools, top_k, top_p | frequency_penalty, include_reasoning, logit_bias, logprobs, max_tokens, min_p, presence_penalty, reasoning, reasoning_effort, repetition_penalty, response_format, seed, stop, structured_outputs, temperature, tool_choice, tools, top_k, top_logprobs, top_p | × differs |
| Defaults | {"temperature":0.6,"top_p":0.95,"frequency_penalty":null} | {} | × differs |
| Reasoning contract | {"mandatory":false} | {"mandatory":false,"default_enabled":true,"supported_efforts":["max","high","low"],"default_effort":"high"} | × differs |
| Modalities | text->text | text+image->text | × differs |
DeepSeek: DeepSeek V3.2 Exp: OpenRouter's listing shows 147,456 output tokens, which is 90% of the context window, the figure it carries where the provider declares no maximum. DeepSeek's own API reference caps max_tokens at 384K, 393,216 tokens, for deepseek-flash and deepseek-v4-pro.
Method: tokenizer fingerprinting, v1, 50 strings. This page is static and dated; the interactive bench compares any two of the 338 measured models.