CEO's Column
Search
More
Foundation Models

India Launches Its Most Powerful Open-Source LLM

ByRishabh Srihari
2025-05-26.about 2 months ago
India Launches Its Most Powerful Open-Source LLM
Sarvam-M, India’s most powerful open-source LLM with 24B parameters, debuts with exceptional performance in Indian languages, math, and code generation.

India’s AI innovation space has welcomed a new contender. Sarvam-M, a 24-billion-parameter large language model (LLM), is the flagship model unveiled by Indian startup Sarvam. Designed for performance and precision, Sarvam-M stands tall among globally recognized models.

A Model Tailored for Deep Understanding

Sarvam-M is built on Mistral Small and enhanced through a meticulous three-stage process. These include Supervised Fine-Tuning (SFT), Reinforcement Learning with Verifiable Rewards (RLVR), and Inference Optimisations.

SFT focused on curating a challenging and diverse dataset. The team filtered generated responses using custom scoring, optimizing for quality and cultural relevance. This step helped Sarvam-M perform well in both complex reasoning tasks and casual conversations.

Through RLVR, Sarvam pushed the model’s boundaries further. By designing reward systems and using specialized prompt sampling, Sarvam-M became highly proficient in logic-heavy tasks like math and programming.

Post-training, the model underwent quantisation to FP8 precision. This improved inference speed without affecting accuracy. The integration of lookahead decoding increased efficiency, though the team notes concurrency issues still need solving.

Also read: Google Cloud Is Reshaping AI Infrastructure to Meet the Demands of the Inference Era

Real-World Applications and Benchmarks

Sarvam-M excels in tasks ranging from conversational AI to machine translation and education. Its strength in Indian language comprehension is a standout feature.

One of its benchmark successes is in the romanised Indian language GSM-8K, where the model achieved an 86% performance improvement. Sarvam-M even outperformed Llama-4 Scout on multiple benchmarks and is comparable to much larger models like Llama-3.3 70B and Gemma 3 27B.

In English-language tasks such as MMLU, it showed a slight 1% dip. Despite this, its performance remains impressive across diverse use cases.

Open-Source Access and Developer Use

Developers can now access Sarvam-M via Sarvam’s API. It’s also available on Hugging Face for direct experimentation and deployment. This availability makes it easy to explore Sarvam-M's potential in real-world projects.

By combining power, precision, and accessibility, Sarvam-M represents a significant leap in India’s AI journey.

Related Topics

Large Language Models (LLMs)LLMs

Subscribe to NG.ai News for real-time AI insights, personalized updates, and expert analysis—delivered straight to your inbox.