Try the Jev Model by TypeSafe in a playground or via API
A 4B model that estimates numbers from text with calibrated intervals
Stage
Startup
Not given
Founded
2026
Not given
Based in
Not given
Not given
Pricing model
Freemium
Freemium
Pricing
Free: 20 welcome credits for new accounts. Token packs: 15M input tokens at $0.600/1M ($9 one-time), 75M tokens at $0.520/1M ($39 one-time), 200M tokens at $0.380/1M ($76 one-time), 600M tokens at $0.283/1M ($170 one-time). Annual subscriptions available at 50% discount. Output tokens are free.
First 1,000 estimates free. Pricing for additional estimates available upon request.
Key features
Yes/No probability questions
Choice questions with confidence
Score questions with rating levels
Browser-based playground
RESTful API with TypeSafe-compatible format
Batch processing tools
Support for images and video (Jev-Omni variant)
Multiple language support
Calibrated confidence scores
Parallel question processing
Returns median, quantiles, and calibrated intervals for numerical estimates
Processes text records in a single forward pass without text generation
Responds in approximately 160 milliseconds
Flags when targets are outside training distribution or intervals lack calibration guarantees
Supports few-shot learning with at least eight labeled examples
Trained on 36 public sources
Calculates probability of exceeding user-specified thresholds
For count targets can return probability of each whole number
Accepts text input without requiring schema declaration or preprocessing
Costs roughly two orders of magnitude less per prediction than frontier models
What makes it different
System One model optimized for structured decisions, not text generation
Calibrated probabilities and confidence scores built into every response type
Fast responses at approximately 0.5 seconds
Independently operated by Jev AI, separate from TypeSafe
Output tokens are free, billing only for input
Provides calibrated uncertainty quantification where language models cannot
Transparently reports when calibration guarantees are absent
Dramatically faster inference than text-generating alternatives at 160ms
Designed for unstructured text without requiring labeled training data
Explicitly signals new, unseen targets with fallback capability for code
Jev AI
maap