
Compact release

Personality-driven language models
JOSIE is a family of language models by Gökdeniz Gülmez—built to pair technical capability with a distinct point of view.
Three scales, one new generation.

Compact release

Preview release

Expanded release
Real words from people trying the new JOSIE generation in the wild.
Each release carries the work forward: new training choices, clearer behavior, and more capable local models.
A personality-driven model family trained entirely on Apple Silicon, spanning 2B, 4B, and 9B releases in practical local formats.
The next iteration of the uncensored, high-performance JOSIE family, available in dedicated Thinking and Instruct variants.
The original JOSIE release established the family’s voice: direct, high-performance, and available across practical local formats.
Version 2
The JOSIE Constitution describes the stable character, values, judgment, and practical habits that animate JOSIE-2: useful, intellectually honest, ethically serious, and capable of treating adults like adults.
Its purpose is to support a coherent character that generalizes to unfamiliar situations.
Download the complete JOSIE Constitution, including the preamble, model lineage, and all sixteen operational articles.
JOSIE explores a more expressive path for small language models: technically serious, personality-forward, and built for people who want to run intelligence on their own terms.
A recognizable voice is treated as a capability—not a decorative layer.
Practical model sizes and local formats keep experimentation close to the user.
Every generation is a visible step in an ongoing, open research process.
Our goal is for more people to run capable, private models on their own hardware—on their own terms.
Prompts, context, and experiments can stay with the person using the model instead of becoming another remote data trail.
Right-sized local models can reduce unnecessary data movement and let people choose efficient hardware for the task at hand. The energy impact still depends on the model, device, and how it is used.
Local formats give people and communities a durable option: useful intelligence that remains accessible, inspectable, and under their control.

JOSIE, local AI tools, and open research are made independently. Sponsorship helps keep the work public, practical, and moving forward.
Sponsor on GitHub