Sarvam AI, a rising player in India’s artificial intelligence landscape, has officially launched its flagship large language model—Sarvam-M. Built on top of Mistral Small, Sarvam-M is a 24-billion-parameter hybrid model with open weights, designed to tackle a wide spectrum of tasks including mathematics, programming, multilingual processing, and Indian language reasoning.
According to the company, Sarvam-M represents a significant advancement in conversational AI, machine translation, and educational applications. It’s already making headlines for setting new standards in math problem-solving, code generation, and Indian language benchmarks.
A Three-Phase Training Framework
The development of Sarvam-M involved a comprehensive three-phase training pipeline:
- Supervised Fine-Tuning (SFT):
Sarvam curated a robust dataset of diverse, high-quality prompts and generated completions using approved models. These outputs were scored, filtered, and fine-tuned to optimize cultural alignment and reduce bias. The model was trained to operate in both “think” mode for complex reasoning and “non-think” mode for general conversation. - Reinforcement Learning with Verifiable Rewards (RLVR):
This phase incorporated programming tasks, mathematical reasoning, and instruction-following datasets. Through custom reward engineering and prompt sampling, Sarvam-M was further optimized to follow instructions accurately and reason effectively. - Inference Optimization:
The team used FP8 quantization to reduce computational overhead with minimal loss in model accuracy. Innovations like lookahead decoding improved throughput, though challenges with higher concurrency were noted.
Benchmark Performance
Sarvam-M has demonstrated outstanding results in key benchmarks:
- On Indian language reasoning tasks like the romanised GSM-8K, it achieved a +86% improvement.
- It outperformed Llama-4 Scout and was found to be on par with much larger models like Llama 3.3 70B and Gemma 3 27B.
- On English knowledge benchmarks such as MMLU, it showed a modest drop of around 1%, which the team considers a reasonable trade-off given its multilingual and reasoning strengths.
Accessibility and Applications
Sarvam-M is available for developers and researchers via Sarvam’s official API and can also be downloaded from Hugging Face for further experimentation and integration. The model is expected to empower a wide array of applications ranging from education and enterprise solutions to local language assistants and voice-based interfaces.
This release comes as part of Sarvam AI’s broader goal to build India’s first sovereign LLM—a step toward democratizing cutting-edge AI technology while tailoring it to local needs and linguistic diversity.
