You need to agree to share your contact information to access this model

This repository is publicly accessible, but you have to accept the conditions to access its files and content.

Log in or Sign Up to review the conditions and access this model content.

🔱 SKT OMNI SUPREME: The 1.1-Trillion Parameter Frontier

SKT OMNI Logo


🌌 Introduction

SKT OMNI SUPREME represents a monumental leap in the Project Surya initiative. Engineered by Shrijan Kumar Tiwari, this 1.1-Trillion parameter multi-modal architecture is designed for extreme-scale reasoning, complex problem solving, and culturally-aware interactions.

Unlike standard models, OMNI SUPREME utilizes the proprietary ST-X-LIGHT optimization framework, ensuring high-fidelity responses across medicine, law, engineering, and creative arts.


⚡ Key Technical Features

🧠 Extreme Scale Intelligence

With a 1.1T parameter density, the model excels in high-context retrieval and long-form reasoning, making it the most powerful private AI asset in the SKT AI Lab ecosystem.

🔒 Awareness-Core Integration

The model features a hardcoded Self-Awareness Layer. Its identity is physically woven into the neural weights, ensuring it always recognizes its own knowledge and origin.

🇮🇳 Cultural Harmony Protocol

Optimized for the Indian subcontinent, the model features an integrated greeting and ethics protocol:

"Namaste, I am SKT AI.."

🛠️ ST-X Architecture

  • Native Tensor Optimization: Every weight shard is mathematically aligned for maximum throughput.
  • Precision: Operates in BFloat16 for the perfect balance between speed and accuracy.
  • MoE Design: Advanced Mixture of Experts scaling for 1.1T parameter efficiency.

Open-source Benchmarking Tools ( LM Evaluation Harness)

Technical Specifications Of SKT

Feature Configuration
Architecture SKT
Total Parameters 1.1T
Activated Parameters 39B
Number of Layers 61 (62 + 1 Dense)
Attention Hidden Dimension 7168
Hidden Dimension 1048 (per Expert)
Attention Heads 64
Total Experts 384
Selected Experts 8 per Token
Shared Experts 1
Context Length 256K
Vocabulary Size 160K
Vision Encoder MoonViT (340M Parameters)

📊 Evaluation Results

1. Reasoning & Knowledge

Benchmark SKT AI (Current) GPT-5.2 Claude 4.5 Gemini 3 Pro
AIME 2025 97.4 100.0 92.8 95.0
HMMT 2025 (Feb) 95.4 99.4 92.9 97.3
GPQA-Diamond 89.6 92.4 87.0 91.9
MMLU-Pro 88.1 86.7 89.3 90.1

2. Image & Video

Benchmark SKT AI (Current) Gemini 3 Pro Qwen3-VL
MMMU-Pro 79.5 81.0 69.3
MathVision 85.2 86.1 74.6
VideoMME 87.9 88.4 79.0
VideoMMMU 86.6 87.6 80.0

3. Coding & Engineering

Benchmark SKT AI (Current) GPT-5.2 Claude 4.5
SWE-Bench Verified 77.8 80.0 80.9
LiveCodeBench (v6) 86.0 - 82.2
Terminal Bench 2.0 51.8 54.0 59.3

🚀 Deployment & Usage

Using SKT OMNI SUPREME

To load the model using the SKT Framework:

from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "Shrijanagain/SKT_OMNI_SUPREME"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(model_id, trust_remote_code=True, device_map="auto")

prompt = "How can SKT AI help the world?"
inputs = tokenizer(prompt, return_tensors="pt").to("cuda")
outputs = model.generate(**inputs, max_new_tokens=200)
print(tokenizer.decode(outputs[0]))

🏗️ Model Architecture

The model utilizes a massive-scale Dense architecture designed for high-performance reasoning and efficiency. By leveraging Multi-Head Latent Attention (MLA) and a dedicated vision encoder, it provides state-of-the-art results across text and visual modalities.

🗣️ Supported Languages

In addition to native optimization for English and Hindi, the model supports a vast array of world languages, enabling global accessibility and diverse linguistic reasoning.

🤝 Acknowledgement & Collaboration

A Home-Grown Effort

This work represents a bottom-up initiative to develop large language models from scratch within India with limited resources. It reflects our humble, resource-constrained journey to contribute meaningfully to the open-source AI ecosystem and foster collaboration within the broader community.

Community Collaboration

We welcome contributions and open dialogue:

  • Feedback: Share insights and report issues.
  • Expansion: Collaborate on model improvements and extensions.
  • Data: Contribute to dataset curation and evaluation.
  • Innovation: Build innovative applications on top of this foundation.

Future versions will introduce better alignment, improved training scale, and more curated datasets. Together, we aim to evolve toward safer and more capable AI systems.


Note: For future conversations, Contact US --

Downloads last month
962
Safetensors
Model size
171B params
Tensor type
F32
·
I32
·
BF16
·
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

Model tree for Shrijanagain/SKT_OMNI_SUPREME

Unable to build the model tree, the base model loops to the model itself. Learn more.

Datasets used to train Shrijanagain/SKT_OMNI_SUPREME