Skip to main content
Decision modelOpen source

Jeff 1

Gestalt Labs

Released
-
Data date
September 24, 2026
Model class
Typed-decision LoRA adapter
Result
Choice, Noul and Score with labels and probabilities
Access
Local

Jeff 1 is a LoRA adapter on a Qwen3-4B-Instruct-2507 base model that scores candidate labels through the language model’s token probabilities. The adapter requires separately loaded base weights. Jeff runs locally, accepts text or structured data, and returns labels and probabilities for Choice, Noul, and Score questions.

A Jev-compatible interface describes technical compatibility, not an affiliation with TypeSafe AI or identical results to Jev. Jeff is an independent project with an Apache-2.0 license for the adapter and source code. You supply the controlled label sets at call time.

Specifications and access

SpecificationValue and source
Model class
Typed-decision LoRA adapterSource
Access
LocalSource
Model ID
GestaltLabs/Jeff-1Source
Input
Text or structured data and typed questionsSource
Output
Structured probabilities, no text generationSource
Result
Choice, Noul and Score with labels and probabilitiesSource
Parameters
4 billion base modelSource
License
Apache-2.0 for the adapter and source codeSource
Scope
LoRA adapter on Qwen3-4B-Instruct-2507 for local inference.Source
Notes
The adapter does not include the base-model weights and is not affiliated with TypeSafe AI.Source
Base model
Qwen/Qwen3-4B-Instruct-2507Source
Adapter
LoRA, rank 16, alpha 32, dropout 0.05Source