01Q4 2025
COMPLETED
L1 Prompt Fusion
Prompts are intelligently generated using ReasonAug and ARD, and prepared for the L2 layer using a specially trained TanAIEncoder.
05 / THE ROADMAP
A system in evolution.
TanAI’s roadmap. Watch and marvel at the evolution of this architecture. It is not necessary to complete all nine layers to use the architecture: each layer is modular and ready for use.
02Q1 2026
REVISION 1-2 COMPLETED
L2 AR Core
Design and coding of the Auto-Regressive core. FUSED, ECV, and CHRONOS Projections completed in the TanAIGATCore architecture.
03Q1 2026
COMPLETED
Model Training (R1-R2)
Training with Corpus data containing millions of lines using the 64K Vocab TanAITokenizer, 1280DIM TanAIEncoder, Offline Indexer, Vector Fused, Emotion, and Time training. Combining the entire architecture to train the Base Model. Our model is currently in the training phase for R1.
04Q2 2026
COMPLETED
Model Training (R3)
The model was scaled up to 2B parameters. R&D processes for the 0.8B model were completed. The tokenizer was retrained with a 96K vocabulary. It supports morphological language structures at lower frequencies. The encoder was trained with a 1536D encoder and a maximum sequence length of 4096. Massive datasets have been indexed for Base-LM training, and the dataset has been vectorized for Condition Projection training. Base-LM is currently being trained.
05Q3 2026
LIVE NOW
New Model Training (R4)
Following the 2B Base and SFT training phases, we are training the ku2-3B model a more advanced model using over 193 billion tokens. We implemented the Transformer Block Attention Residual enhancement and resolved the RoPE and KV-truncation issues encountered in the 2B model. This model will serve as TanAI's foundational model.
06Next Step
QUEUED
L3 VAE / Bridge
We will design a bridge that combines the inferences of the AR model with dLLM.
07Next Step
QUEUED
L4 dLLM Core
The second core of the architecture is the Diffusion architecture. The backbone of the Memory-Prediction structure. It will process latent representations from L2 and L3 to produce humanoid and contextually sensitive outputs.
08Next Step
QUEUED
dLLM Training
The hardest part of the job is training the Diffusion Core by creating noise with the datasets in the AR model.
09Next Step
QUEUED
L5 Fusion Core Layer
We will design our layer that transforms the multi-dimensional outputs of AR and dLLM into a single, coherent semantic-emotional representation.
10Next Step
QUEUED
L6 Emotional Memory
The layer that ensures the user's emotional and semantic continuity and guarantees the contextual consistency of the generated output.
11Next Step
QUEUED
L7 Out Synthesis
The final point of inference from all layers. The decoder of Voice Profiler plugins in its most stable form.
12Next Step
QUEUED
L8 Self Adaptive Tuning
SAT refines what it learns from the user through a pipeline structure and reflects it in the Model's weights. It reduces the embedded RAG load.
13Next Step
QUEUED
L9 Agentic Planner
They apply what they've learned without needing you that's all there is to it...