Indian AI startup Sarvam has unveiled its flagship open-source Large Language Model (LLM), Sarvam-M, marking a major leap in the country’s journey toward sovereign AI development. Featuring an impressive 24 billion parameters, Sarvam-M is tailored for high-performance across a variety of tasks including mathematics, programming, and Indian language processing.
Built upon Mistral Small, the model has been fine-tuned to support critical AI applications such as conversational AI, machine translation, and educational tools. Its training methodology incorporates Supervised Fine-Tuning (SFT), Reinforcement Learning with Verifiable Rewards (RLVR), and inference-level optimizations to enhance performance and cultural relevance.
During the SFT phase, Sarvam focused on generating high-quality prompts and filtering model responses to maintain cultural appropriateness, especially in the Indian context. The innovative RLVR technique employed custom reward engineering over datasets specific to instructions, mathematics, and coding, ensuring the model’s outputs are both accurate and contextually aligned.
The company further boosted the model’s efficiency through post-training enhancements, including FP8 quantization and lookahead decoding, enabling faster and more resource-efficient inference without compromising quality.
In terms of performance, Sarvam-M delivered exceptional results, particularly in Indian languages and mathematical reasoning, where it demonstrated an 86% improvement on benchmarks like GSM-8K. The model not only outperformed Llama-4 Scout but also showed competitive performance against top-tier models like Llama-3.3 70B and Gemma 3 27B, with just a marginal (~1%) drop in English benchmarks.
With this release, Sarvam positions itself at the forefront of India’s AI ecosystem, promoting open-source innovation while reducing dependence on foreign models. Sarvam-M reflects a growing trend toward creating localized AI systems capable of addressing unique linguistic and cultural nuances, and is expected to power a wide range of India-centric digital applications in the near future.
