Every few weeks, someone in Indian AI claims the next big thing. This week it's actually worth a look. CoRover.ai — the Delhi-based company behind the BharatGPT family of sovereign AI models — has published fresh benchmark results for BharatGPT-Instruct, a model with roughly 2 billion active parameters. And the short version is: it beat Sarvam 30B, a model around fifteen times its size, on nearly every reasoning and fairness benchmark it was tested on.
If you've been following this space, you'll remember we wrote about Sarvam's Indus chat app and, more recently, Gnani's Artha sovereign stack. India's homegrown AI race is getting genuinely competitive — and now, interestingly, it's getting smaller.
01.The numbers behind the BharatGPT win
Let's get straight to the scores, because they're the story here. On five standard benchmarks, BharatGPT-Instruct came out ahead of both Qwen 1.7B and Sarvam 30B:
- ARC-Easy: 83.21% — a grade-school science reasoning test
- ARC-Challenge: 53.84% — the harder version of the same test
- WinoGrande: 68.67% — common-sense word understanding
- PIQA: 78.51% — physical commonsense reasoning
- CrowS-Pairs: 68.28% — a fairness benchmark that measures social bias
A 2B model out-reasoning a 30B model on ARC-Challenge, and doing better on a bias benchmark too — that's not a rounding error. That's a design philosophy winning. As CoRover founder and CEO Ankush Sabharwal put it: "Scale is not the same as intelligence... sovereign AI is not about building the biggest model; it is about building models that are efficient, relevant, trustworthy."
02.Why a smaller model winning matters for your wallet
Here's the thing nobody's talking about: model size is basically a running cost. A 2B model needs far less compute to serve, which means cheaper inference, lower latency, and the ability to run on modest hardware — think a startup's server rather than a GPU cluster. For Indian businesses watching every rupee, that's the difference between a demo and an actual deployment.
It's the same logic we keep hammering in our GPT-6 Astra — Zero to Hero Guide, which has an entire chapter on cost optimization — most people overpay for AI because they use more model than the job needs. Matching the smallest capable model to the task is the single biggest cost lever you have, whether you're building on BharatGPT, Sarvam or GPT-6 Astra.
03.The 10-language test BharatGPT aced
The more impressive part for me is the multilingual result. On Multilingual MMLU, BharatGPT-Instruct led Qwen 1.7B across all ten regional languages tested: Hindi, Bengali, Marathi, Telugu, Gujarati, Malayalam, Punjabi, Tamil, Odia and Kannada. The model was built for India first, and it shows — this isn't an English model with a Hindi paint job.
04.The RAG result that should get enterprises excited
For companies, the benchmark that actually matters is RAG — retrieval-augmented generation, where an AI answers using your own documents instead of its training data. BharatGPT-Instruct scored 100% on faithfulness, ahead of GPT-4o-mini (95%) and Qwen 3 1.7B (95.51%). In plain terms: when it answers from your knowledge base, it stays honest to what's actually in there. CoRover also shared its governance practices — PII filtering, anonymisation, and a no-training, deployment-only policy for sensitive and regulated clients.
05.Credit where due: they published the bad score too
Honestly, what surprised me most wasn't the wins — it was the loss they volunteered. On GSM8K, the maths reasoning benchmark, BharatGPT-Instruct scored just 23.50%, against 68% for Qwen 1.7B and 70.60% for Sarvam 30B. Publishing that gap openly is rare in this industry, where every launch is "best-in-class" at something. CoRover says it did this deliberately so future development targets real weaknesses. Whether that's marketing or maturity, it's the right instinct.
06.Context: CoRover is not a small player
Worth knowing who's behind this. CoRover's platform claims to have reached over 1.8 billion lives, with 65 million-plus monthly active users and 95,000-plus enterprises, developers and researchers, across 100+ languages. It already runs PortGPT, a domain-specific model for the shipping ports sector, with more verticals in the works. The BharatGPT-vs-Sarvam benchmark dust-up is really a preview of the next phase of India's sovereign AI push — efficiency over brute force.
Key Takeaways
- BharatGPT-Instruct (~2B parameters) beat Sarvam 30B on ARC, WinoGrande, PIQA and CrowS-Pairs benchmarks
- It led Qwen 1.7B on Multilingual MMLU in all ten Indian languages tested
- 100% RAG faithfulness — ahead of GPT-4o-mini — makes it serious for enterprise deployments
- Openly weak on maths: 23.50% on GSM8K, and CoRover published it anyway
- Smaller models mean cheaper inference — the cost argument matters as much as the scores
Frequently Asked Questions
What is BharatGPT-Instruct and what makes it notable?
CoRover.ai's BharatGPT-Instruct has roughly 2 billion active parameters yet outscored Sarvam 30B, a model about 15 times its size. It won on ARC-Easy (83.21%), ARC-Challenge (53.84%), WinoGrande (68.67%), PIQA (78.51%) and CrowS-Pairs (68.28%).
Where does BharatGPT struggle?
It beat Qwen 1.7B on Multilingual MMLU across all ten Indian languages tested and scored 100% on RAG faithfulness, ahead of GPT-4o-mini — but it openly struggles on maths (GSM8K: 23.50%).
What This Means for You
If you're a developer or a startup in India, stop assuming you need the biggest model for everything. A well-trained 2B model that understands ten Indian languages and stays faithful to your documents can power a customer-support bot, a vernacular content tool, or a government-facing service at a fraction of the running cost of a 30B giant. If you're a student, this is your signal that understanding model evaluation and cost trade-offs is a real, hireable skill. And if you want to actually learn that skill properly — which models to pick, how to prompt them, how to keep costs sane — the GPT-6 Astra Zero to Hero Guide is the practical next step; it walks through all of it with copy-paste templates you can apply to any model, BharatGPT included.