AI Desk · AI Fingerprinting

Every model, in one place

Name: YFarmX AI Fingerprinting catalogue
Creator: YFarmX

Reference desks for every kind of model: language, image, video, audio, world models, agents and robotics. Underneath them sits the YFarmX lab, which measures the tokenizer behind each endpoint, checks the serving contract, and keeps every observation as dated history.

Try:

Models by type

Start from what the model does; each desk carries the reference pages

Large Language ModelsThe models that read, reason and write, from the frontier labs to the open-weight field.90 pagesOpen
Image GenerationText to image and image editing models, from frontier quality to the open-weight stack.19 pagesOpen
Video GenerationText and image to video models, from the safe Western defaults to the leaderboard leaders.21 pagesOpen
Audio and VoiceMusic, speech and real-time voice models, and the licensing questions they raise.11 pagesOpen
World ModelsSystems that simulate interactive, explorable environments a frame at a time.10 pagesOpen
Agentic AIThe autonomous agents that run tools, write code and act on your behalf, not just answer.18 pagesOpen
RoboticsMachines that act in the world: the models that drive them, the bodies they drive, and the rules around both.14 pagesOpen

In-house lab · YFarmX

AI Fingerprinting

We measure what the label leaves out

A model's page tells you what its maker calls it. We send fifty adversarial strings to the endpoint itself and read the token counts that come back, which is the tokenizer speaking rather than the marketing. That is how this desk named two anonymous models before the companies behind them did.

337models measured on 50 strings69distinct fingerprints56tokenizer families found2stealth models named first

Models measured: 337
Complete fingerprints: 314
Tokenizer groups: 56
Unlabelled models placed: 103 of 118
New since 16 Sept: 31

Catalogue and measurements as at 8 October 2026. Measured counts every model with a validated measurement; complete fingerprints are the subset whose 50 rows all came back clean, and only those carry an exact signature. Every figure derives from one committed dataset; the methodology records how it is measured, and what the method cannot see.

The latest identification

Named from the measurements 18 hours before its operator confirmed it

Union Alpha

stealth/union-alpha · Circuit and ChiselConfirmed by its maker

Confirmed family: Pareto (multi-model blend)
Confirmed as: Pareto 26.9, built by Circuit and Chisel
Confirmed by The Unbiased Co: 17 September 2026
Entered catalogue: 16 September 2026
Last tested: 17 September 2026
Family attribution confidence: Very high

Why very high: the operator confirmed Union Alpha as Pareto 26.9 on 17 September 2026, 18 hours after the YFarmX post named it; the blend carries the Llama 3, Qwen and GLM tokenizers the measurements had read.

Evidence

✓ near-exact agreement (46/47) with meta-llama/llama-3.2-3b-instruct
✓ confirmed as Pareto 26.9, built by Circuit and Chisel by The Unbiased Co, 17 September 2026

Compare with Meta: Llama 3.2 3B Instruct Full identity record

What the measurements say

Findings the current dataset supports

Confirmed callOx Alpha matched Z.ai: GLM 5.3 on all fifty strings, and Z.ai confirmed itA stealth endpoint with no maker named returned token counts identical to GLM 5.3 on every one of the fifty test strings, sharing the signature tk_133b086c. YFarmX published the match on 21 August 2026. Z.ai confirmed the model as GLM-5.3-Flash five days later.See the evidence →Measured twice66 models measured twice: 61 fingerprints held, 3 movedOn 5 October 2026 the most-used models on record were sent the fifty strings again and read against their earlier reading. 61 agreed on every string clean in both runs. 3 did not, and each of those records says which strings moved, by how much, and which provider served each reading.See the evidence →ResolvedSpace Bunny Alpha: a stealth model with a measured familySpace Bunny Alpha declares no family. Independent measurement places it with the MiniMax group at high confidence: exact tokenizer signature shared with 8 models.See the evidence →

Explore the lab

Six ways into the same measurements

01Every modelAll 337 measured models on one page, each row opening its evidence.Open the catalogue →02CompareTwo models side by side: tokenizer agreement string by string, API surface field by field.Open the bench →03Identity watchNew arrivals, fingerprint changes and fresh lineage placements as they are observed.See what changed →04UnknownsStealth endpoints and the 118 models with no family on their label. 103 placed so far.Open the queue →05Tokenizer AtlasThe family structure of the whole catalogue: clusters, constellation and the explorer.Open the Atlas →06MethodologyHow a fingerprint is measured, the controls that gate publication, and the known limits.Read the method →

How it works

01 · MEASURE

Fifty adversarial strings are sent to each model and only the reported prompt-token count is read. The marginal cost of each string under the model's tokenizer becomes a 50-number signature.

02 · COMPARE

Signatures are compared offline against every other model measured. Exact agreement across all fifty strings is strong lineage evidence; the API-surface contract is checked as a second signal of a different kind.

03 · TRACK

Every observation is kept as taken, and the history only ever appends. When an endpoint's fingerprint moves, the history shows what changed and on which day.

A shared tokenizer is evidence of shared lineage, never proof of an exact model. The claims hierarchy keeps every published conclusion inside what the evidence supports, and the controls are published with the results.

Run a fingerprint comparison →Pick any two of the 337 measured models and see all fifty strings