India’s Answer to DeepSeek: The Race to Launch Indigenous LLMs by End-2025
Updated: 6 days ago
By Parikshit Khanna — Premium AI Strategy | Delivered with Authority, Not ApologiesJuly 2025 | 1,050 words
AI news update: week of Thursday 12 November 2026
This article was first published earlier. Here are the main AI news items of this week, each with its source.
Anthropic released two new models, Claude Fable 5.2 and Claude Mythos 5.2, on Monday 9 November 2026. Reports say they are built for coding and multi step agent work. Source: AI Agent Store weekly news.
Anthropic announced on Tuesday 10 November 2026 that it plans to raise up to 64.4 billion US dollars in a share sale, according to Axios. Source: Axios.
Google DeepMind said its AI system AlphaProof Code checked the security of the Signal messaging app's encryption code using machine checked proofs. Source: The Neuron daily digest.
Microsoft began fair use limits for Copilot Chat users who do not have a Microsoft 365 Copilot licence. The rollout started on Monday 2 November 2026 for smaller organisations and on Monday 9 November 2026 for larger ones. Source: Microsoft Partner Center announcements.
OpenAI said its Advanced and Standard Voice modes will be retired on Friday 11 December 2026, and users will be moved to Live Voice. Source: The Neuron daily digest.
In India, the Andhra Pradesh Cabinet approved the Robotics and Physical AI Policy 2026 to 2036 on Tuesday 10 November 2026. Source: Elets CIO.
China’s DeepSeek shocked the global AI scene in January 2025 with its DeepSeek-R1 model — open weights, ChatGPT-class performance, sub-$10M cost. It wasn’t just a model. It was a geopolitical statement.
India took the message seriously.
As of July 2025, India isn’t theorizing about foundational AI anymore. It’s building one. Actually, multiple. And if you’ve been asleep at the wheel, this post is your overdue wake-up call.
Let’s dissect India’s LLM roadmap: the actors, the infrastructure, the timelines — and what serious professionals need to do now.
1. The Trigger: DeepSeek-R1’s Release
When DeepSeek launched in early 2025, it proved that:
You don’t need OpenAI’s $100M+ budgets
You can deliver LLM parity with just a few billion tokens of tightly curated training data
Sovereign countries can build performance-grade models with strategic constraint, not infinite scale
Indian policymakers, already pushing DPI (Digital Public Infrastructure), saw this as validation. Less than 60 days later, India’s ₹10,372 crore IndiaAI Mission was restructured to explicitly prioritize sovereign LLMs.
2. The Build: Compute, Data & Model Proposals
India’s LLM launch plan is not a headline. It’s infrastructure-backed and funded.
Here’s what’s already in motion:
10,000+ H100/MI300-class GPUs procured under central infrastructure policy
TGDeX, India’s first government-led AI dataset exchange, is live in Telangana
43 of 506 model proposals approved under IndiaAI explicitly target foundation models or LLM fine-tunes
GPU subsidies slashing inference costs to ₹100/hr or lower
First wave of multilingual datasets is open-sourced under the Bharat DataSet Consortium
You don’t launch a DeepSeek competitor overnight. But you do it faster when you control policy, compute, and national-scale training data.
3. The Players: Sarvam AI and BharatGen
Let’s name names.
📍 Sarvam-M (Sarvam AI)
Released May 2025
24B parameter multilingual LLM
Trained on India-heavy corpora — code, maths, science, and over 10 Indic languages
Funded through the IndiaAI Mission
Performance: Promising in low-resource use-cases, but not yet at DeepSeek-R1’s level
Sarvam AI is first to market. But it’s also a test case — to validate alignment, voice input, and regulatory integration.
📍 BharatGen (Govt-backed LLM Consortium)
Led by IITs, CDAC, Jio R&D, and select cloud players
Native multimodal support — text, audio, image
22 languages targeted in alpha; script-level tokenisation for faster fine-tuning
Dataset mix includes DigiYatra, Ayushman Bharat logs, and PM-Kisan inputs under anonymised compliance
Release timeline: Alpha rollout in Q3, with full model ready by Q4 2025
This is India’s DeepSeek moment, planned for Diwali window. If you’re in AI ops, this is your countdown clock.
4. The Policy: India’s Strategic AI Stack
No serious LLM program works without policy firepower. India is using carrots + safeguards:
40% capital subsidy on GPUs
Fast-track regulatory sandbox for AI in health, agri, and fintech
Encrypted non-local data storage allowed — enabling AWS/OpenAI usage under India’s key-hold policies
Trusted AI Audits framework in draft, to govern models above 1B parameters
That’s what separates hobby models from DeepSeek-grade deployments: a regulated runway, not wild-west scaling.
5. Timeline: When India’s DeepSeek-Class LLMs Land
Quarter | Key Milestone |
Q3 2025 | Sarvam-M iteration 2 releases with improved benchmark scores. BharatGen Alpha opens to developers via DPI credentials |
Q4 2025 | BharatGen launches with DeepSeek-parity architecture + multilingual evaluation suite |
Q1 2026 | First gov use-cases (e.g., Ayushman claim processors, AI-powered Helplines) go live in 3 states |
Mid-2026 | Public-private APIs built atop BharatGen and Sarvam for Indian enterprise scale |
The mission is clear: India wants domestic AI horsepower before the 2026 Lok Sabha session reconvenes.
6. Implications for Strategy, Hiring & Tech Stack
If you’re leading a product, ops, AI, or innovation role in India — this isn’t FYI material. This is plan-of-record input.
Here’s what you do now:
✅ Audit Your LLM Dependencies
Are you hard-locked to OpenAI or Anthropic APIs?
Have you tested compatibility with Sarvam/BharatGen alpha releases?
✅ Prepare for Data Transfer
TGDeX APIs are going live. If you’ve been sitting on India-specific structured/unstructured data — plug it into model adaptation workflows now.
✅ Pre-Train Your Teams
Sarvam and BharatGen will launch APIs, doc portals, and plugin support. Don’t wait for production-grade; train your devs on the alpha.
✅ Shift Your UX
BharatGen is voice-native and multimodal. That means your front-ends, mobile flows, and retrieval layers must evolve.
✅ Comply Early
If you're deploying any LLM in healthcare, government, or BFSI — your compliance teams must review the Trusted AI Audit drafts right now.
7. Bottom Line: India’s LLM Push Is Execution, Not Optics
We’re past the “can we do this?” phase. India’s funding is in, the GPUs are delivered, the data exchanges are live, and the first models are running.
Will BharatGen outperform DeepSeek on Day 1? Maybe not.Will it be adopted nationally before DeepSeek ever gets legal clearance in India? Absolutely.
That’s the bet. And it's working.
🔐 Final Word: Act Now, Or Be Outpaced
The launch of DeepSeek was China’s masterstroke. India’s answer — Sarvam-M, BharatGen, and DPI-linked sovereign LLMs — is strategic, policy-anchored, and weeks away from prime-time.
Don’t wait for perfect benchmarks or Gartner reports.
This is the moment to embed, align, and deploy.Or you’ll spend 2026 explaining to your board why your stack runs on deprecated APIs from another continent.
Books by Parikshit Khanna
PDF books on Topmate. Prices are as listed on Topmate on Thursday 12 November 2026.
Master AI Prompting: 3,000 Prompts for Real Work
3,000 ready prompts for everyday work tasks. ₹99 on Topmate, earlier listed at ₹999.
The Agentic AI Playbook 2026
A 735 page guide to AI agents at work. ₹99 on Topmate.
The AI Workbook: Practical Fill-In Templates
Fill-in templates to apply AI to your own tasks. ₹99 on Topmate.
Photos from Parikshit Khanna's training sessions
Small team working on laptops around a meeting table during a corporate AI session with Parikshit Khanna Parikshit Khanna speaking on stage at TEDxEicher School Faridabad Youth, August 2026 Group photo from the Generative AI and ChatGPT training at the National Academy of Broadcasting and Multimedia, Delhi, with Parikshit Khanna Managers working on laptops during a hands on corporate AI session with Parikshit Khanna Parikshit Khanna on the TEDxEicher School Faridabad Youth stage giving his talk Redesigning Work with Artificial Intelligence Participants working on laptops at a shared table during a corporate AI training session





