Sarvam-M: A “Thinking” Large Language Model Empowering Indian Languages

What is Sarvam-M?

Sarvam-M is an open-source large language model (LLM) developed by Indian AI startup Sarvam AI. Based on the Mistral Small architecture (24B parameters), it has been deeply fine-tuned and optimized for India’s multilingual environment. The model supports two distinct inference modes: “non-think” and “think,” adapting to different task complexities.
It is designed to enhance AI’s ability to understand and generate content in Indian languages, driving the vision of “Sovereign AI” in India.

Key Features

Multilingual Support: Optimized for major Indian languages, including Hindi, Tamil, Telugu, Malayalam, Odia, Gujarati, Kannada, Bengali, Marathi, and Punjabi.
Reasoning & Coding Capabilities: Performs exceptionally well on tasks like math (GSM-8K) and coding (HumanEval, MBPP), with up to +86% improvement on Indian language math benchmarks.
Dual Inference Modes: Supports both “non-think” and “think” modes to handle tasks of varying complexity.
Efficient Deployment: Achieves high throughput and low latency using FP8 quantization and TensorRT-LLM compiler optimizations.

Technical Overview

1. Supervised Fine-Tuning (SFT)

Sarvam-M leverages a curated set of diverse prompts, ranked by quality and difficulty, to train a high-quality prompt-completion dataset. This enables debiased and rebias training and supports hybrid dual-mode reasoning.

2. Reinforcement Learning with Verifiable Rewards (RLVR)

For tasks like instruction following, math, and coding, the model is trained using reward models and the GRPO algorithm to optimize for reliable task completion.

3. Inference Optimization

Post-training quantization (PTQ) is used to produce an FP8 model. Inference is accelerated using lookahead decoding strategies and performance tuning on H100 GPUs.

Project Links & Access

Application Scenarios

Multilingual Customer Support Bots: Works across channels like phone and WhatsApp to improve service efficiency.
Legal Document Generation: Assists lawyers in drafting and reviewing contracts and agreements.
Education & Training: Provides AI-powered tools for multilingual teaching, especially in math and programming.
Government & Public Services: Enhances accessibility to e-governance services in local languages.