You can just train models

Your task is specific. Your model should be too.

Join the 19th beta and turn your examples into small, fast models for focused product tasks, built to beat frontier LLMs on accuracy, latency, and cost.

Make a model. From your data. In minutes.

19th lets anyone train their own models from scratch. ML team optional.

Extraction

PII Identifier

Small on-device PII identifier model allows redaction before data upload.

Built from 14,000 structured sequences

Scoring

Real-time Content Toxicity

On-device sub-10ms scoring of content toxicity across 6 discrete categories.

Built from 160,000 user comments

Scoring

Grocery Add-on Recommendations

Instant add-on recommendations based on current cart, recent items, and store availability.

Built from 3M ordered carts

Classification

Live Support Routing

Given a customer enquiry, this model routes to the correct department in 13ms.

Built from 13,000 customer interactions

Learned Scoring

Learned Lead Scoring

Looking back over what actually converted to derive optimal lead scoring, MQL, and where to automate outreach.

Built from 9,000 leads

Scoring

Ad Campaign Planner

Recommended campaign configuration, spend, and ad units based on advertiser profile, target geo, and budget.

Built from 9,100 ad campaigns

Classification

Tic Tac Toe

When models become effectively free to train we can start using them for silly things.

Built from 36,000 synthetic moves

Extraction

PII Identifier

Small on-device PII identifier model allows redaction before data upload.

Built from 14,000 structured sequences

Scoring

Real-time Content Toxicity

On-device sub-10ms scoring of content toxicity across 6 discrete categories.

Built from 160,000 user comments

Scoring

Grocery Add-on Recommendations

Instant add-on recommendations based on current cart, recent items, and store availability.

Built from 3M ordered carts

Classification

Live Support Routing

Given a customer enquiry, this model routes to the correct department in 13ms.

Built from 13,000 customer interactions

Learned Scoring

Learned Lead Scoring

Looking back over what actually converted to derive optimal lead scoring, MQL, and where to automate outreach.

Built from 9,000 leads

Scoring

Ad Campaign Planner

Recommended campaign configuration, spend, and ad units based on advertiser profile, target geo, and budget.

Built from 9,100 ad campaigns

Classification

Tic Tac Toe

When models become effectively free to train we can start using them for silly things.

Built from 36,000 synthetic moves

Extraction

PII Identifier

Small on-device PII identifier model allows redaction before data upload.

Built from 14,000 structured sequences

Scoring

Real-time Content Toxicity

On-device sub-10ms scoring of content toxicity across 6 discrete categories.

Built from 160,000 user comments

Scoring

Grocery Add-on Recommendations

Instant add-on recommendations based on current cart, recent items, and store availability.

Built from 3M ordered carts

Classification

Live Support Routing

Given a customer enquiry, this model routes to the correct department in 13ms.

Built from 13,000 customer interactions

Learned Scoring

Learned Lead Scoring

Looking back over what actually converted to derive optimal lead scoring, MQL, and where to automate outreach.

Built from 9,000 leads

Scoring

Ad Campaign Planner

Recommended campaign configuration, spend, and ad units based on advertiser profile, target geo, and budget.

Built from 9,100 ad campaigns

Classification

Tic Tac Toe

When models become effectively free to train we can start using them for silly things.

Built from 36,000 synthetic moves

Extraction

PII Identifier

Small on-device PII identifier model allows redaction before data upload.

Built from 14,000 structured sequences

Scoring

Real-time Content Toxicity

On-device sub-10ms scoring of content toxicity across 6 discrete categories.

Built from 160,000 user comments

Scoring

Grocery Add-on Recommendations

Instant add-on recommendations based on current cart, recent items, and store availability.

Built from 3M ordered carts

Classification

Live Support Routing

Given a customer enquiry, this model routes to the correct department in 13ms.

Built from 13,000 customer interactions

Learned Scoring

Learned Lead Scoring

Looking back over what actually converted to derive optimal lead scoring, MQL, and where to automate outreach.

Built from 9,000 leads

Scoring

Ad Campaign Planner

Recommended campaign configuration, spend, and ad units based on advertiser profile, target geo, and budget.

Built from 9,100 ad campaigns

Classification

Tic Tac Toe

When models become effectively free to train we can start using them for silly things.

Built from 36,000 synthetic moves

Extraction

PII Identifier

Small on-device PII identifier model allows redaction before data upload.

Built from 14,000 structured sequences

Scoring

Real-time Content Toxicity

On-device sub-10ms scoring of content toxicity across 6 discrete categories.

Built from 160,000 user comments

Scoring

Grocery Add-on Recommendations

Instant add-on recommendations based on current cart, recent items, and store availability.

Built from 3M ordered carts

Classification

Live Support Routing

Given a customer enquiry, this model routes to the correct department in 13ms.

Built from 13,000 customer interactions

Learned Scoring

Learned Lead Scoring

Looking back over what actually converted to derive optimal lead scoring, MQL, and where to automate outreach.

Built from 9,000 leads

Scoring

Ad Campaign Planner

Recommended campaign configuration, spend, and ad units based on advertiser profile, target geo, and budget.

Built from 9,100 ad campaigns

Classification

Tic Tac Toe

When models become effectively free to train we can start using them for silly things.

Built from 36,000 synthetic moves

Purpose-built vs Frontier LLM. What happens when you use the right tool for the job.

All models tested against Banking 77 classification task

Speed

160x faster

12.6ms

Compared to Opus 4.6: 2016ms

Cost per thousand tasks

310x cheaper

$0.03

Compared to Opus 4.6: $9.25

More accurate

+23.4%

94.6% accuracy

Compared to Haiku 4.5: 77.5%

19th lets you build models that extract classify score rank transform

Hundreds of candidates.
One production model.

The most laborious elements of ML. Automated.

19th searches hundreds of model candidates — across architectures, sizes, and training strategies — and distills the best into a production model. What an ML team does in weeks, finished in minutes.

Search

Architectures, sizes, training strategies -> each generation informed by the last.

Distill

Top performers teach smaller, faster deployment models.

Ship

Production-ready. Passes robustness, latency, and cost thresholds.

Train your first task-specific model.

Bring your task and the examples you already have. 19th will handle the rest.