Update README.md

README.md

@@ -4,6 +4,41 @@ language:
 - en
 tags:
 - agent
+- ternary
+- 1.5-bit
+- overflow
+- large-scale
+- efficiency
 ---
 
-# Coming Soon.. (:<
+# 🌊 Overflow-1T
+
+**Overflow-1T** is a next-generation, **1.03 Trillion parameter** Large Language Model built on a custom **1.5-bit Ternary ({-1, 0, 1}) architecture**. 
+
+By utilizing the **Overflow** architecture, this model achieves massive scale reasoning while remaining computationally efficient, designed specifically to run on consumer-grade hardware through advanced weight packing and specialized C++ inference kernels.
+
+## 🚀 Key Specifications
+* **Parameters:** 1,000,000,000,000 (1T)
+* **Precision:** 1.5-bit Ternary (packed 5-weights-per-byte)
+* **Architecture:** OverflowForCausalLM
+* **Layers:** 128
+* **Hidden Size:** 16,384
+* **Attention:** Grouped Query Attention (GQA) with 16 KV heads
+* **Format:** `.safetensors` / `.bbuf` (Optimized for 1TSumerGPU)
+
+
+
+## 🛠 Project Status: Initial Sharding
+We are currently in the process of sharding the 1.5-bit weights to the Hugging Face Hub. 
+- **Progress:** Shard 1 of 10 currently uploading.
+- **Estimated Completion:** March 2026.
+
+## 🧠 Why 1.5-bit?
+Unlike standard 1-bit models, Overflow-1T utilizes a **0-state** (Neutral weight). This allows the model to effectively "silence" noise across its 1T parameter space, leading to significantly higher stability in Chain-of-Thought (CoT) reasoning and logic tasks compared to binary 1-bit models.
+
+## 💻 Inference
+This model is designed to be served using the **1TSumerGPU** engine, a custom C++ and CUDA-based inference framework optimized for NVIDIA RTX 40-series GPUs.
+
+---
+**Created by CooLLaMACEO**
+*Part of the Kwen Foundation initiatives.*