TURKISH-NATIVE. CORE-FIRST.

Intelligence,
from the
core.

We build language models from the inside out. A Turkish-native foundation, a modular architecture, and a research journey you can follow.

IN DEVELOPMENTTanAI 3B Ku2
01 / GLASSBOX CORE+
FUSEDECVCHRONOSGAT / CONDITION STACKFIG. 001

Semantic continuity, carried through the core.

CURRENT MODEL
3B Ku2
TRAINING CORPUS
193B+ tokens
RESEARCH FRAMEWORK
9 modular layers
BUILT FOR
Turkish TR

01 / OUR APPROACH

Beyond the next token.
Towards deeper context.

TanAI explores how meaning, emotion and time can become part of a model’s internal representations. The Condition Stack brings these dimensions into the Generative Adaptive Transformer.

01

FUSED

Meaning in continuity.

Semantic context carried through the model’s conditioning vectors.

02

ECV

A sense of emotion.

Emotional Conditioning Vectors that represent the nuances of an input.

03

CHRONOS

Context includes time.

Temporal projections for relationships between events and their timing.

An evolving architecture, with a visible development roadmap.Explore the nine layers ↗

02 / THE FOUNDATION MODEL

TanAI 3B Ku2.

Built from scratch. Trained for Turkish. Our fourth research revision brings a 3-billion-parameter model to an 8×H100 training cluster.

Meet the model ↗

Training in progress · Project status, September 2026

R0
153.2M
R1
182.7M
R2
813.16M
R3
2.058B
R4
3B
PARAMETERS / R0 → R4

03 / FROM THE LAB

Research, in writing.

All publications ↗

04 / RESEARCH IN MOTION

The latest developments.

All updates ↗

05 / A TRACEABLE JOURNEY

Follow the work.

From the first Prompt Fusion experiments to the current training run. Explore the decisions, milestones and work still ahead.

Where we’re goingRoadmap ↗How we got hereDevelopment log ↗