Sarvam-M: India’s 24 Billion-Parameter LLM Is Challenging Global Leaders

• Chinmay
• May 29, 2025
• Artificial Intelligence, India, News
• 24B parameter LLM, AI in India, AI startups India, GSM-8K Indian benchmark, Indian language AI, Indian LLM, Llama vs Sarvam, LLM benchmarks, Mistral Small, open source language model, open weight models India, RLHF, RLVR, Sarvam AI, Sarvam-M
• 0 Comments

Indian AI startup Sarvam has taken a bold step in the global AI race with the release of its flagship large language model, Sarvam-M — a 24-billion parameter open-weights hybrid LLM that’s setting new performance benchmarks in math, programming, and Indian language understanding.

Built on top of Mistral Small, Sarvam-M is more than just another LLM. It’s been carefully engineered through a three-stage pipeline — Supervised Fine-Tuning (SFT), Reinforcement Learning with Verifiable Rewards (RLVR), and Inference Optimization — to balance complex reasoning and general-purpose conversational capabilities.

What’s Special About Sarvam-M?

Sarvam-M stands out not just because of its size, but because of its versatility and cultural grounding. From machine translation to educational tools, conversational AI, and math/code reasoning, the model is designed for wide adoption.

• SFT focused on quality prompts and curated completions, aiming to reduce cultural bias and teach the model to “think” and “chat”.
• RLVR further trained Sarvam-M on programming, logic, and math through custom reward engineering and prompt sampling strategies.
• Inference Optimizations included post-training quantisation to FP8 and lookahead decoding to boost performance with negligible accuracy loss.

How It Performs

Sarvam-M doesn’t just compete — it wins across multiple metrics:

• Achieved +86% improvement on GSM-8K benchmark for romanised Indian language + math tasks.
• Outperforms Llama-4 Scout on most benchmarks.
• Comparable to larger models like Llama 3.3 70B and Gemma 3 27B.
• Slight (~1%) dip in English-heavy benchmarks like MMLU.

In short, Sarvam-M proves that size isn’t everything — strategy, cultural context, and engineering finesse matter just as much.

Why This Matters

India has entered the LLM arena with a model that doesn’t just speak English well, but understands local languages, educational needs, and real-world constraints. For developers, researchers, and startups seeking powerful open-source AI tools grounded in Indian realities, Sarvam-M is a landmark moment. As the world races to build smarter models, Sarvam-M shows that India’s not just catching up — it’s setting new directions.