AI Desk · Model Identity · Compare

Nous: Hermes 4 70B · Sao10K: Llama 3 8B Lunaris

Identity similarity: lowDivergent fingerprints · 12/50These models tokenise differently. Divergence on the high-information strings (CJK, emoji, digits) separates unrelated vocabularies within a handful of samples.

Nous: Hermes 4 70B recordSao10K: Llama 3 8B Lunaris recordOpen in the interactive bench

Tokenizer fingerprint

12 of 50 mutually clean strings agree
Test stringNous: Hermes 4 70BSao10K: Llama 3 8B LunarisVerdict
en-prose1413differs
en-long1615differs
spaces-2032differs
spaces-6033match
tabs-2032differs
newlines-2044match
mixed-ws54differs
digits-932differs
digits-1243differs
digits-301010match
digits-sep76differs
float-long98differs
zh-common1211differs
zh-long1818match
zh-rare1818match
ja-kana1110differs
ja-kanji1211differs
ko1413differs
ru1413differs
ar1211differs
he2423differs
hi1514differs
th1414match
el1312differs
emoji-basic109differs
emoji-skin3029differs
emoji-zwj-family1515match
emoji-zwj-x34544differs
emoji-flags2424match
emoji-prof2625differs
math3130differs
boxdraw2322differs
combining1110differs
cjk-ext-b1313match
surrogates4039differs
zalgo3534differs
rtl-mix87differs
py-code2524differs
py-indent1514differs
json2019differs
html1615differs
regex4444match
camel54differs
snake66match
rare-word-x53129differs
repeat-tok2220differs
base643534differs
hex1111match
url1716differs
uuid2726differs

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
FieldNous: Hermes 4 70BSao10K: Llama 3 8B LunarisVerdict
Context length1310728192× differs
Max output tokens16384× differs
Supported parametersfrequency_penalty, include_reasoning, max_tokens, presence_penalty, reasoning, repetition_penalty, response_format, temperature, top_k, top_pfrequency_penalty, logit_bias, logprobs, max_tokens, min_p, presence_penalty, repetition_penalty, response_format, seed, stop, structured_outputs, temperature, top_k, top_logprobs, top_p× differs
Defaults{}{}✓ match
Reasoning contract{"mandatory":false}× differs
Modalitiestext->texttext->text✓ match

Method: tokenizer fingerprinting, v1, 50 strings. Raw data on the data page. This page is static and dated; the interactive bench compares any two of the 422 catalogue models.