Update README.md

README.md

@@ -1,4 +1,32 @@
 ---
 license: cc0-1.0
 ---
-![](img/pre.png)
+![](img/pre.png)
+
+# VDrontV3-Mini - custom MoE model. (gpt 2 tokenizer)
+
+You can find `model.py` in the repository; it describes the model architecture.
+
+The repository also contains `use.py` and `GUIvdront.py`—example scripts for running the model. To use them, download the model files and place them in the `VDrontV3-Mini/` folder, then place `model.py`, `use.py`, and `GUIvdront.py` alongside that folder and run the script.
+
+The model has two modes: "Base" and "Quality." In "Base" mode, responses are shorter and more varied, whereas in "Quality" mode, they are more similar in structure but longer than those in "Base".
+
+# GUIvdront.py
+![](img/chat.png)
+
+# model architecture
+![](img/ar.png)
+
+
+*Base model VDrontV2-0.1b*
+
+VDrontV3 / MeB-LM (Many Expert Block LM) is a transformer-based architecture that implements a hybrid Mixture of Experts (MoE) design with two specialized expert blocks.
+
+    Embedding & Routing: The input passes through standard token and position embeddings. A "Router top-1" (a linear layer) processes the embedding to select a single expert (top-1 routing).
+
+    Expert Block: A sequence of transformer layers is replicated into multiple parallel "Expert Blocks" (purple boxes). The router dynamically activates exactly one of these expert versions during each forward pass.
+
+    Base Block: After the expert block, the hidden states pass through a shared, fixed "Base Block" (green box) consisting of standard transformer layers. This provides a common foundation for all routing paths.
+
+    Output Block & Mode Block: The final layers form an "Output Block" (also multi-version) that feeds into the LM head. Crucially, the architecture includes a "ModeBlock" (yellow box), which suggests the model can switch between different inference modes by selecting different versions of the output block.
+