CoRover.ai Claims Its BharatGPT-Instruct Model Outperforms Sarvam 30B

CoRover.ai has claimed that its BharatGPT-Instruct model, with around 2 billion active parameters, has outperformed Sarvam 30B, a model roughly 15 times its size, on several key AI benchmarks.

BharatGPT-Instruct led Sarvam 30B and Qwen 1.7B across several benchmarks and achieved 100% faithfulness in CoRover’s enterprise RAG evaluation. However, it underperformed on the GSM8K mathematics benchmark.

According to CoRover, the model outperformed both Sarvam 30B and Qwen 1.7B with scores of] 83.21% on the elementary science knowledge benchmark ARC-Easy, 53.84% on the advanced science reasoning benchmark ARC-Challenge, 68.67% on the common-sense reasoning benchmark WinoGrande, 78.51% on the physical solutions evaluation PIQA, and 68.28% on CrowS-Pairs, which evaluates social bias.

The new model challenges the assumption that bigger AI models are necessarily better, shifting the focus to smaller models that prioritise efficiency over scale while performing across local languages. 

“Scale is not the same as intelligence,” Ankush Sabharwal, Founder and CEO of CoRover.ai, said in a statement. According to Sabharwal, when AI is designed around a country’s problems, languages and requirements, model size does not necessarily determine performance.

“It is about building models that are efficient, relevant, trustworthy and capable of solving real-world problems at scale,” he added.

BharatGPT-Instruct also performs strongly across Indian languages, the startup claimed, including] Hindi, Bengali, Marathi, Telugu, Gujarati, Malayalam, Punjabi, Tamil, Odia, and Kannada, surpassing Qwen 1.7B and Sarvam 30B on Multilingual MMLU.

PortGPT, CoRover’s domain-specific version of BharatGPT, designed for the ports and shipping sectors, is already live, while other specialised models are under development.

Earlier this year, the Indian Army’s Department of Military Affairs selected CoRover to set up a sovereign AI-ML lab powered by BharatGPT to enable Army personnel to learn, build, and deploy secure AI agents.

The AI startup is also empanelled by the National e-Governance Division under the Ministry of Electronics and Information Technology as a trusted AI implementation partner for the public sector.