The AI trained exclusively on content that went viral.

Lena
LSMKJB
Generate
Hook
Template
Structure
“Stop ___. Seriously.”
Beats
⏱ 0.0s
⏱ 2.8s
Aa 7words
↯ 1cut
Delivery
T  Deadpan
P  To camera
Maya
Generating hook variantsV1.1.0
Scoring variants against SVBench™…
Rewriting for 2.8s hook window…

Write, score, and ship scroll-stopping scripts directly from the canvas

N
Hooks
Export
Pattern interruptTikTok · 0–3s

“Stop buying cold brew. Seriously. I did the math.”

91
SVBench™ 91.4Top 3% of 1,284 matched viral hooks

No more guessing. No more AI slop. Viral IS a pattern.

A model that has watched more viral content than any human ever could.

ViralML™ is 100% fine-tuned on hand-picked content that actually went viral. Not scraped, not synthetic. Every transcript was selected by our editorial team, annotated for structure, and trained on our own GPUs in Frankfurt am Main.

0
hours of viral video transcripts in the fine-tuning set.
≈ 52.5 years of watch time
100%
hand-picked. Every sample cleared a virality threshold and an editorial review.
0 synthetic samples
0pt
ahead of the next-best frontier model on SVBench™ overall.
FRA
Trained and served from our own datacenter in Frankfurt. Your briefs never leave the EU.
fra-1 · ISO 27001 · GDPR

How we built the viral index

General models learned to write from the whole internet, which is mostly content nobody watched. We did the opposite.

01
Hand-pick the outliers

Only posts that beat their creator's median by 10× or more. Across TikTok, Reels, Shorts, X and LinkedIn.

02
Annotate the anatomy

Hooks, beats, pattern interrupts, open loops and CTAs, labelled second by second across 460,000 hours.

03
Fine-tune in Frankfurt

Full-parameter fine-tuning on our own cluster. No third-party APIs between your brief and the model.

04
Measure against reality

Every checkpoint is scored on SVBench™: real posts, real audiences, blind panels.

AI alone can write. ViralML™ writes what gets watched.

Read the methodology
SVBench™ overall scorev0.9 · preliminary
n = 12,400 held-out briefs · higher is better
ViralML 1.0Ours
91.4
Opus 5.5
68.2
Fable 5.1
64.7
Sonnet 5.5
61.3
Astra 6
57.9
Kimi K3
53.6
0255075100
Pro copywriter median · 74.0
Frontier models evaluated zero-shot with an identical system prompt and brief. Hover a bar for sub-scores.Full leaderboard

Everything a brand needs to be the one people share

“Same list, same reps, same offer. We just stopped writing like everyone else.”

Halden's outbound team switched every sequence to ColdLLM™ for one quarter. Reply rates climbed, meetings got better, and their close rate rose 24%.

JW
Jonas Weber
VP Sales, Halden Systems
Read the case study
Close rate, opportunities → won
+0%
Before11.3%
With ColdLLM™14.0%
2.1×
positive reply rate
−38%
time spent writing sequences

Stop posting into the void.

You're on the list. We'll be in touch from Frankfurt.

Brands already signed up for early access