Tamil-R1: First AI That Actually THINKS in Tamil (Yeah, verified with Google)

Hi everyone! 👋

I'm excited to share Tamil-R1 - the first reasoning model that performs chain-of-thought reasoning natively in Tamil.

The Problem I Solved

I tested ChatGPT o1, DeepSeek R1, Sarvam-M, Claude, and Gemini with Tamil questions. Every single one showed reasoning in English and then translated to Tamil.

Even Sarvam-M (built in India with $41M funding) thinks in English when you click "Show Thinking."

That's not Tamil AI. That's English AI in Tamil clothing.

What Makes Tamil-R1 Different

Other models:

Tamil question → English thinking → Tamil answer

Tamil-R1:

Tamil question → Tamil thinking → Tamil answer

Example:

Question:

காய்கறி வாங்கணும். பக்கத்து கடை ₹45, தூர கடை ₹35. எது நல்லது?
(Need to buy vegetables. Nearby shop ₹45, distant shop ₹35. Which is better?)

Tamil-R1 Thinking:

<think>
இரண்டு வழி பார்ப்போம்:

பக்கத்து கடை:
- விலை = ₹45
- நேரம் = 5 நிமிடம்
- வசதியானது

தூர கடை:
- விலை = ₹35
- நேரம் = 30 நிமிடம்
- கஷ்டம்

நேர மதிப்பு கணக்கு:
- ₹10 சேமிப்புக்கு 25 நிமிடம் கூடுதல்
- மணிக்கு ₹24 மதிப்பு மட்டும்
- உன் நேரம் இதுக்கு அதிக மதிப்பு
</think>

Answer:

பக்கத்து கடையே போ. ₹10 சேமிப்புக்கு 25 நிமிடம் பயணம் மதிப்பு இல்ல. 
நேரம் = பணத்தை விட முக்கியம்.

See the difference? Pure Tamil reasoning. No translation artifacts. No English thinking.

How I Built It

• Time: 2 weeks
• Cost: ₹0 (using Kaggle free GPU)
• Dataset: 100 hand-crafted Tamil reasoning examples
• Method: Fine-tuned DeepSeek-R1-Distill-Llama-8B using LoRA
• Training: 30 minutes on Tesla T4 x2
• License: MIT (100% open source)

Technical Details

• Base Model: deepseek-ai/DeepSeek-R1-Distill-Llama-8B
• Fine-tuning Method: LoRA (rank 16, alpha 16)
• Training Data: 100 Tamil examples across 5 categories:
  • Shopping & Money decisions
  • Home & Cooking planning
  • Daily decisions
  • Health & Routine
  • Common sense reasoning
• Training Loss: 2.73 → 2.07 (3 epochs)
• Format: User question → <think>Tamil reasoning</think> → Answer

Limitations (v0.1)

⚠️ This is a proof-of-concept with only 100 training examples:

• May generate repetitive or incomplete outputs
• Limited to everyday scenarios (not complex academic problems)
• Needs 5-10x more training data for production use
• No standard benchmark evaluation yet

This is v0.1 - rough around the edges, but it WORKS.

Why This Matters

70 million Tamil speakers deserve AI that:

• ✅ Thinks in their language
• ✅ Understands cultural context
• ✅ Reasons naturally without translation artifacts
• ✅ Shows transparent thinking process

Language shapes thought. When AI thinks in English and translates, it loses:

• Cultural nuance
• Natural reasoning patterns
• Linguistic authenticity

Try It Yourself

Note: This is my first model.