Moondream: Moondream 3.1

moondream/moondream3.1-9b-a2b · MoondreamTokenizer lineage measured

Tokenizer group (measured): Moondream
Entered catalogue: 15 October 2026
Last tested: 17 October 2026
Evidence confidence: Very high

Why very high: an exact fingerprint shared with 1 model and an identical API contract inside the same group: two signals of different kinds agree.

Evidence

✓ exact tokenizer signature shared with 1 model
✓ API-surface contract identical to a group member

Compare with Moondream: Moondream 3 Preview

Report a correction

Identity statement

Moondream: Moondream 3.1 is Moondream's model, and its measured tokenizer sits in Moondream's own signature group alongside its sibling models. No other maker's measured model shares it so far.

What we checked

Four different kinds of evidence, and what each one showed

We checked	What we found	Where it came from
How it counts tokens	Counts every one of the 50 test strings exactly like 1 other model	tk_2442c19b
How its API is set up	Set up identically to 1 other listing, field for field	api_1a80e003
How much it can read and write	Reads up to 32,000 tokens at once, writes up to 4,096	its own listing
Whether it thinks before answering	Reasons when asked to	its own listing
Who served our requests	PrimeIntellect	we measured it, 17 October 2026

Shares this fingerprint

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

Moondream: Moondream 3 Previewmoondream/moondream3-previewexact · 50/50

Closest measured models

Agreement across the mutually clean test strings

Moondream: Moondream 3 Previewmoondream/moondream3-preview50/50 · 100%IBM: Granite 4.0 Microibm-granite/granite-4.0-h-micro15/50 · 30%IBM: Granite 4.1 8Bibm-granite/granite-4.1-8b15/50 · 30%Microsoft: Phi 4microsoft/phi-415/50 · 30%OpenAI: GPT-3.5 Turboopenai/gpt-3.5-turbo15/50 · 30%OpenAI: GPT-3.5 Turbo 16kopenai/gpt-3.5-turbo-16k15/50 · 30%

Click a row to open the full comparison.

The measured fingerprint

What each test string cost this model, in tokens

We sent Moondream: Moondream 3.1 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.

All 50 rows

English and whitespace

en-prose12

en-long15

spaces-203

spaces-606

tabs-207

newlines-2021

mixed-ws8

Digits

digits-93

digits-124

digits-3010

digits-sep7

float-long9

CJK

zh-common30

zh-long49

zh-rare26

ja-kana17

ja-kanji26

ko49

Other scripts

ru20

ar30

he30

hi26

th52

el25

Emoji

emoji-basic15

emoji-skin30

emoji-zwj-family18

emoji-zwj-x354

emoji-flags24

emoji-prof30

Rare Unicode

math32

boxdraw32

combining14

cjk-ext-b16

surrogates41

zalgo36

rtl-mix12

Code

py-code16

py-indent12

json22

html18

regex53

camel8

snake8

Repetition and encodings

rare-word-x531

repeat-tok41

base6439

hex12

url21

uuid27

Overhead subtracted: 5 prompt tokens.

Declared record

What the catalogue claims about this model

Moondream 3.1 is a vision-language model from Moondream, the successor to Moondream 3 Preview, with 9B total parameters and 2B active (mixture-of-experts). It handles visual question answering, captioning, OCR, and...

Modalities: text+image->text
Declared tokenizer: Other
Prompt price: $0.30 / M tokens
Completion price: $2.50 / M tokens

Supported parameters

include_reasoningmax_completion_tokensmax_tokensreasoningtemperaturetop_p

Default settings

Temperature: 0.2
Top p: 0.9

Catalogue entry Weights on Hugging Face

History

Every observation, kept as taken

15 October 2026Enters the OpenRouter catalogue with no declared family.
17 October 2026Fingerprinted in the YFarmX catalogue sweep · 50 of 50 strings measured clean.
17 October 2026YFarmX assessment: consistent with the Moondream family, very high confidence.