alethic
independent research · Uzbekistan

Small models, measured honestly.

Alethic is language-model architecture research with the work in the open: weights, data, and a terminal agent you can download and run yourself. English, Uzbek and code.

alethic — ~/portfolio
> add a dark mode toggle to the site
reading layout.tsx, styles.css …
<map> layout.tsx (Main layout) · styles.css (Theme) · hero.tsx (Hero)
editing styles.css lines 40–72
✓ done, 2 files changed
>
Illustrative session. Not a recording.
300M
parameters in Alethic-300M
EN · UZ · code
what the model is trained for
open
weights on Hugging Face
1 binary
terminal agent, no account

Everything is here, nothing is hidden.

Run the agent, load the weights, read the data, check the numbers.

agent

The Alethic agent

A terminal UI with tools: read, write, shell and search. Context stays small because the agent maps sections of what it has read instead of carrying it all. Orchestrators can spawn sub-agents for bigger jobs.

models

From 1M to 300M, all public

Each model ships with its config and a model card. Pull them from Hugging Face.

See the models →
datasets

Alethic-newborn

13,810 instruction examples in English and Uzbek, with tool-use, thinking and long project sessions. Access is gated: accept the terms, then download.

Read the dataset card →
research

Matched baselines, negative results included

Sin-Gated Control Attention is tested against a conventional GQA transformer at matched parameters and compute. The write-ups say what worked and what did not.

Read the research log →

One person, built on constrained compute.

Alethic is led by Bekhruz Suleyman, an independent researcher and engineer from Uzbekistan. The method is simple: understand the mechanism, build the smallest real version, measure it, find what is wrong, repeat. Every experiment is designed to be reproducible on free or low-cost compute.

start here
Run Alethic in your terminal in a few minutes.