Sarvam-M: India's AI model impresses in maths and Indian languages; here's how it compares to other AI models

Indian AI startup Sarvam has launched its flagship large language model (LLM), Sarvam-M, a 24-billion-parameter hybrid open-weights model built on Mistral Small. Positioned as a versatile, locally relevant alternative in the global LLM race, Sarvam-M has earned praise for its strong performance in Indian languages, mathematics, and programming, but not without some scepticism from parts of the tech community.

24 billion parameters: What does it mean?

In simple terms, parameters are the internal settings a language model uses to process and generate text. Think of them as dials and switches that get tuned during training to help the model understand grammar, context, facts, reasoning, and more. The more parameters a model has, the more nuanced its understanding and output can be, though this also depends on data quality and training methods. Sarvam-M, with 24 billion parameters, falls into the mid-to-large scale of LLMs. It’s significantly larger than open models like Mistral 7B, but smaller than frontier systems like OpenAI’s GPT-4 or Google’s Gemini 1.5 Pro.

How does Sarvam-M stack up?

Here’s a quick look at where Sarvam-M fits among major players:

Model	Parameters	Strengths
Sarvam-M	24B	Indian languages, maths, programming
OpenAI GPT-4	1.8T (estimated)	General reasoning, coding, multilingual
Gemini 1.5 Pro	200B+	Multimodal capabilities, advanced reasoning and coding performance
Llama 3 70B	70B	Reasoning, coding, and multilingual tasks
Anthropic Claude 3.7 Sonnet	2T (estimated)	High-quality summarisation, reasoning, and content generation

Sarvam-M sits below the largest proprietary models in terms of size but punches above its weight in domain-specific tasks, particularly maths and Indian language reasoning. However, it trails behind in English-centric benchmarks such as MMLU, with about a 1% performance gap, highlighting room for improvement in broader linguistic generalisation.

How was it built?

Sarvam-M was developed through a three-phase training process:

1. Supervised Fine-Tuning (SFT): High-quality prompts and responses were used to shape the model’s conversational and reasoning abilities while minimising cultural bias.
2. Reinforcement Learning with Verifiable Rewards (RLVR): The model was taught to follow instructions and solve logic-heavy problems using carefully engineered rewards and feedback loops.
3. Inference Optimisation: Advanced compression (FP8 quantisation) and decoding strategies helped improve efficiency and speed, though scalability challenges in high-concurrency environments persist.

Why it matters

Sarvam-M supports 10 Indian languages and can handle competitive exam questions in Hindi, making it a promising tool for local education and translation efforts. It achieved an 86% improvement in a test combining maths and romanised Indian languages, demonstrating strong multilingual reasoning.

Despite criticism over whether the model is “good enough” to compete globally, Sarvam-M’s launch has significantly raised the profile of Indian efforts in the AI space. The model is now publicly accessible via Sarvam’s API and on Hugging Face, encouraging developers to build, test, and contribute.

While it may not rival the most advanced LLMs just yet, Sarvam-M represents a meaningful step forward for democratising AI development in India, especially for users who need support beyond English.