Sarvam AI, an Indian Artificial Intelligence company, has launched its flagship language model, Sarvam-M, a powerful open-source LLM featuring 24 billion parameters. Built on top of the Mistral Small architecture, Sarvam-M combines advanced capabilities in mathematics, programming, and Indian language processing, positioning itself as a versatile model for a wide array of AI-driven applications.
The model is designed to support various real-world use cases, including conversational agents, machine translation systems, and educational tools, because of its hybrid architecture and refined training process. Sarvam-M can tackle complex reasoning challenges across disciplines such as logic, code generation, and multilingual comprehension.
The team assembled a carefully curated dataset of high-quality prompts to fine-tune the model. By generating responses using vetted base models and applying custom filters, they were able to minimise cultural bias while maximising relevance. The SFT approach trained Sarvam-M to operate effectively in both high-level reasoning scenarios (“think mode”) and more straightforward dialogue settings.
For efficient deployment, Sarvam-M was optimised post-training through quantisation to FP8 precision, achieving faster inference times with minimal impact on output accuracy. Additional enhancements like lookahead decoding were introduced to improve processing speed, although the developers noted scalability challenges related to concurrent usage.
Sarvam-M’s reasoning performance was further improved using a specialised training curriculum. This phase included tasks such as following instructions, solving math problems, and writing code. Customised reward functions and smart prompt selection techniques were employed to sharpen the model’s abilities across these domains.
Sarvam-M has demonstrated exceptional performance in mixed-language benchmarks, particularly those combining Indian languages with mathematical reasoning. For example, in the romanised GSM-8K benchmark tailored to Indian linguistic patterns, the model delivered an 86% improvement over its base counterpart.
In broader evaluations, Sarvam-M has outperformed smaller LLMs like Llama-4 Scout and shows competitive results against larger models such as Meta’s Llama 3.3 70B and Google’s Gemma 3 27B. However, it lags slightly, by approximately 1%, on certain English-focused tests like MMLU.
Developers and researchers can now access Sarvam-M via the company’s official API. The model is also freely available for download on Hugging Face, allowing seamless experimentation, integration, and further community-driven innovation.
With Sarvam-M, Sarvam AI aims to advance the development of AI solutions rooted in India’s linguistic and educational diversity, contributing to a more inclusive global AI ecosystem.
Be a part of Elets Collaborative Initiatives. Join Us for Upcoming Events and explore business opportunities. Like us on Facebook , connect with us on LinkedIn and follow us on Twitter.
"Exciting news! Elets technomedia is now on WhatsApp Channels Subscribe today by clicking the link and stay updated with the latest insights!" Click here!
