StepFun: Step 3.5 Flash

stepfun/step-3.5-flash · StepFunTokenizer lineage measured
Tokenizer lineage (measured)
DeepSeek
Evidence confidence
High
Why high: an exact fingerprint shared with 9 models that declare the same family, without a second signal of a different kind yet.
Declared tokenizer tag
Other
Entered catalogue
29 January 2026
Last tested
21 August 2026
Measurement
Measured
Evidence
  • exact tokenizer signature shared with 9 models

Identity statement

StepFun: Step 3.5 Flash is StepFun's model. Its measured tokenizer sits with the DeepSeek group, which is evidence that it builds on the DeepSeek tokenizer. Tokenizer reuse is normal engineering, open vocabularies travel between labs, and this says nothing against StepFun's authorship of the model itself.

Verification note

Checked against the primary artefacts, 21 August 2026

StepFun's own Step-3 repository ships a tokenizer behaviourally identical to DeepSeek-V3's: run over the 50 strings offline the two agree on every one, and both declare a 128,815-entry vocabulary. The endpoint's counts match as well, so this model sitting in the DeepSeek signature group is genuine lineage reuse by StepFun, not a serving artefact.

StepFun step3 repository (config.json: vocab_size 128815)DeepSeek-V3 repository (config.json: vocab_size 129280; tokenizer table shared)

Evidence stack

Signals of different kinds, weighed together
SignalResultReference
Tokenizer signatureExact match with 9 other modelstk_73c4f0dc
API surfaceNo other entry shares this exact contractapi_b9b93634
Context and output262,144 context · 65,536 max outputdeclared
Reasoning contractMandatorydeclared
Serving providers observedSiliconFlowmeasured 21 August 2026

Shares this fingerprint

Exact signature first; template-boundary shifts of the same signature beneath

Closest measured models

Agreement across the mutually clean test strings

Click a row to open the full comparison.

The measured fingerprint

Marginal prompt-token cost of each test string, grouped by script
All 50 rows

English and whitespace

en-prose14
en-long16
spaces-203
spaces-603
tabs-204
newlines-204
mixed-ws7

Digits

digits-93
digits-124
digits-3010
digits-sep7
float-long9

CJK

zh-common9
zh-long10
zh-rare18
ja-kana9
ja-kanji11
ko13

Other scripts

ru13
ar10
he14
hi18
th11
el16

Emoji

emoji-basic10
emoji-skin14
emoji-zwj-family11
emoji-zwj-x333
emoji-flags16
emoji-prof19

Rare Unicode

math22
boxdraw27
combining10
cjk-ext-b16
surrogates33
zalgo36
rtl-mix6

Code

py-code25
py-indent15
json20
html16
regex47
camel7
snake8

Repetition and encodings

rare-word-x526
repeat-tok41
base6434
hex12
url17
uuid27

An x marks a row excluded as corrupt (caching or provider interference detected during measurement). Overhead subtracted: 12 prompt tokens.

Declared record

What the catalogue claims about this model

Step 3.5 Flash is StepFun's most capable open-source foundation model. Built on a sparse Mixture of Experts (MoE) architecture, it selectively activates only 11B of its 196B parameters per token....

Modalities
text->text
Declared tokenizer
Other
Prompt price
$0.10 / M tokens
Completion price
$0.30 / M tokens

Supported parameters

frequency_penaltyinclude_reasoningmax_tokensreasoningtemperaturetool_choicetoolstop_ktop_p

defaults: {"temperature":null,"top_p":null,"top_k":null,"frequency_penalty":null,"presence_penalty":null,"repetition_penalty":null}

Catalogue entryWeights on Hugging Face

History

Every observation, kept as taken
  • 29 January 2026Enters the OpenRouter catalogue with no declared family.
  • 21 August 2026Fingerprinted in the YFarmX catalogue sweep · 50 of 50 strings measured clean.
  • 21 August 2026YFarmX assessment: consistent with the DeepSeek family, high confidence.

Cite this page as the evidence record for StepFun: Step 3.5 Flash: the URL is stable, measurements are dated, and revisions append to the history above. Method: tokenizer fingerprinting, v1, 50 strings. The raw data behind every figure is on the data page.