Skinnertopia Lab for AI
Independent open model researchOpen models should be able to reason
SLAI builds open models that can reason, build, work, design, explore, and collaborate, along with the products that make those abilities useful.
Latest from the lab
Journal
Research notes, model work, product releases, and direct accounts of what we are learning.
Why SLAI Exists
Local AI gave people ownership of a model, but ownership was not enough. SLAI began with a harder question: when would open intelligence become dependable work?
Read article
Controlling More Than Thought
Reasoning effort controls how deeply a model deliberates. Work effort controls how far an agent carries the task. Useful systems need both dials.
Read article
The Work Begins When the Prompt Ends
A useful agent does not race from prompt to answer. It explores the project, tests its understanding, acts, and checks what actually changed.
Read articleOur premise
Democratizing more than intelligence.
Publishing weights is a beginning. People also need the controls, interfaces, tools, and workflows that turn open intelligence into dependable work.
Open model families
Models built for every task.
GRaPE and CRePE are developed for people who need more than an answer. They are designed to carry work through.
General Reasoning Agent
GRaPE
General Reasoning Agent for Project Exploration. A multimodal family for reasoning, coding, tools, and complex project work.
Explore GRaPE
Code Reasoning Expert
CRePE
Code Reasoning Expert for Project Exploration. Focused on agentic software work with separate control over thought and work effort.
Explore CRePEProducts
The model is not the finish line.
SLAI builds the surfaces that let open models become useful parts of real projects.
Conversation and collaboration
GRaPE Chat
A hosted workspace for SLAI models, files, tools, artifacts, visible reasoning, and collaborative Vines.
Open GRaPE Chat
Agentic coding
Scribe
A terminal agent built around project exploration, deliberate work, and first-class support for GRaPE and CRePE models.
Install Scribe
Research
Control the thought. Control the work.
Useful agents need control over both their reasoning depth and the amount of work they carry through.
Read our researchControlling More Than Thought
Reasoning effort controls how deeply a model deliberates. Work effort controls how far an agent carries the task. Useful systems need both dials.
Read articleSince August 2025
Local intelligence was open, but it could not yet do the work.
SLAI began after years of watching local models promise ownership without delivering dependable agency. Qwen 2.5 and Qwen 3 showed that the gap could close.
GRaPE 1 established the direction. GRaPE 2 caught up with the competition. The lab exists to keep moving forward and to publish the products each model family deserves.