ੴ Sehaj 30M: The Sovereign World-Ready AI Model
☬ ਸਹਿਜ 30M — ਪੰਜਾਬੀ, ਹਿੰਦੀ ਅਤੇ ਅੰਗਰੇਜ਼ੀ ਦਾ ਸੁਤੰਤਰ ਨਿਊਰਲ ਮਾਡਲ
Created and Architected by Gurpreet Singh Dhillon (Toon Studio).
Sehaj 30M is an ultra-efficient, byte-level native autoregressive neural architecture (29.6M parameters) designed for 100% on-device local execution on edge devices and mobile phones.
🌟 Major Breakthrough Highlights / ਨਵੀਆਂ ਵਿਸ਼ੇਸ਼ਤਾਵਾਂ
- Q&A Instruction SFT (0.149 Validation BPB): Fine-tuned on Gurpreet Singh's 50,000-sample high-density Gurmukhi & Multilingual instruction corpus.
- Multilingual Tri-Core Engine: Native byte-level processing for Punjabi (Gurmukhi), Hindi (Devanagari), and English (Latin) without token vocabulary explosion.
- Low-Latency Neural Voice (VITS PTTS): Real-time native voice synthesis with high-pass clarity filtering.
- On-Device Vision OCR: Real-time camera reading of books, receipts, medicine labels, and Indic scripts.
- Autonomous Dynamic Tool Engine: On-demand generative micro-apps (Water Tracker, Grocery Ledger, Smart Ring Health Monitor, Food Calorie Scanner).
- Universal Everything Phone Controller: Voice-controlled orchestration across all 100+ installed applications and phone hardware (Flashlight, Dialer, Alarms, Volume).
📊 Model Specifications / ਮਾਡਲ ਜਾਣਕਾਰੀ
- Author / ਰਚੇਤਾ: Gurpreet Singh Dhillon (Nam-toon Studio)
- Parameters / ਪੈਰਾਮੀਟਰ: 29,613,944 (~29.6M)
- Architecture: Causal Byte-Level Transformer with Vismaad Surprisal Gating & Surat Hebbian Memory
- Validation BPB: 0.149 BPB on held-out instruction test set
- APK Package:
SahajAI-30M.apk(55 MB standalone Android application) - License: MIT License (Open Source)
📂 Repository Contents / ਫਾਈਲਾਂ
m30_qa_finetuned.pt: Fine-tuned PyTorch checkpoint (0.149 BPB)SahajAI-30M.apk: Complete ready-to-install Android Applicationsehaj_omni_service.py: FastAPI server combining Chat, TTS Voice, Vision, and Toolssehaj_v32.py: Model architecture definition
🚀 Quick Run (Python / Terminal)
# 1. Clone repo
git clone https://huggingface.co/Nam-toon-studio/sahaj-30m
# 2. Run local Omni-Service
python3 sehaj_omni_service.py
Created with dedication by Gurpreet Singh Dhillon (Toon Studio).
- Downloads last month
- 741