ReasonOS Research

Intelligence that learns from experience.

Smaller, faster models, and models that improve through experience. Measured results, with their methods and limits.

Model efficiency · internal results

Over 70% smaller. About 2× faster.

We make trained models over 70% smaller without conventional quantization. In separate optimization experiments, inference ran about twice as fast.

Model sizeOriginal · 100%Compressed · 30%
Inference speedOriginal · 1×Compressed · ~2×
Separate experiments, not one combined configuration. Internal measurements.

Continual learning · working demo

AI that improves through experience.

Our General Learning Operator updates a model’s weights from the work it completes. In an internal demonstration, a model kept what it learned from earlier coding tasks and did better on the ones that followed.

Memory and retrieval

The model stays the same. Past conversations or documents are looked up and pasted into the prompt every time.

Learning in the weights

The model itself changes. What it learned shapes the next answer without being looked up again.

In the product todayReasonOS’s standing agents keep session state and project context. Their model weights don’t change; that is what this research is for.

Active researchLearning from feedback, recursive improvement, less compute per update, and learning that stays isolated to one person or organization.

Why it matters

Cheaper AI that gets better at your work.

Today’s AI comes from a few large general models: expensive to train, expensive to run, and the same for everyone. Efficient models that learn from use change that.

Cheaper to run

Smaller models and faster inference need less hardware for every request. That changes where AI can run and who can afford it.

Specialized by use

A model that learns from the work it does gets better at that work, instead of staying as general as the day it was trained.

Owned by its user

When learning lives in a model you hold, what it learns from you stays yours, isolated from every other customer.

Where we are headed

Intelligence that belongs to you.

We are working toward specialized models that become more useful to each person or organization over time. Each learns only from the people it serves, and that learning stays isolated from everyone else. This is the direction of the research; it has not shipped.

Results

The technical record.

Both results are internal experiments. We will publish methods and data with our papers.

Model compression

Internal experiment
Objective
Make trained models smaller and faster to serve without losing output quality.
Method
Our own compression method, applied to a trained model, without conventional quantization. Inference speed was measured in separate optimization experiments.
Baseline
The same model, uncompressed.
Results
Over 70% smaller. In the optimization experiments, about 2× faster inference.
Limitations
Measured on our own benchmarks. Size and speed come from separate experiments and are not one combined result. Not yet published or reviewed outside the lab.
Reproducibility
Model, benchmark and methodology are shared with research partners on request.

Learning from prior coding tasks

Internal experiment
Objective
Show that a model can keep useful knowledge from completed work in its weights.
Method
The General Learning Operator updates the model after each completed task. No retrieval and no external memory.
Baseline
The same model without the updates.
Results
The updated model kept what it learned and performed better on the coding tasks that followed.
Limitations
An internal demonstration on coding tasks. Recursive improvement and other domains are active research.
Reproducibility
Shared with research partners on request.

Published writing and talks

People

Who is doing the work.

Vamsi Pandari

Founder

Built the ReasonOS developer platform and the infrastructure under it, including the build system, branch servers and the compute the research runs on.

Background in AI infrastructure and high-performance computing.

LinkedIn

Dr. Ehsan Ardjmand

AI Researcher

Leads the mathematical work behind our model compression and continual-learning research.

PhD. Mathematician and AI researcher.

LinkedIn

Collaborate

Work with us.

We partner with researchers and companies on compression and continual learning. Tell us what you are working on.