Sarvam AI launches flagship LLM, comparable to Meta, Google models
What's the story
Sarvam AI, a Bengaluru-based start-up selected under the Indian government's IndiaAI Mission, has launched its flagship Large Language Model (LLM), Sarvam-M.
The 24-billion-parameter multilingual, hybrid-reasoning text-only language model is based on Mistral Small, an open-weight AI model created by French firm Mistral AI.
The Sarvam-M has been designed to cater to a variety of use cases such as conversational agents, translation services, and educational tools.
Performance
Surpassing benchmarks in Indian languages, math, programming
Sarvam-M has established new benchmarks across different Indian languages, math, and programming tasks.
The AI model exhibits a 20% average improvement over the base model on Indian language benchmarks, a 21.6% enhancement on math-related tasks, and a 17.6% improvement in coding benchmarks.
At the intersection of Indian languages and mathematics, it shows an impressive +86% improvement in romanized Indian language GSM-8K benchmarks.
The current version follows launch of new speech model, Bulbul, which features authentic accents that truly sound Indian.
Comparison
Sarvam-M outperforms LLaMA-4 Scout, comparable to larger models
Sarvam AI claims that the advanced Sarvam-M model outperforms Meta's LLaMA-4 Scout on most benchmarks and is comparable to larger dense models like LLaMA-3.3 70B and Google's Gemma 3 27B.
These models are pre-trained on significantly more tokens.
However, the company acknowledges that there's room for improvement in "knowledge related benchmarks in English," where Sarvam-M drops about 1% point over the baseline model MMLU.
Sarvam-M is open source and available on Hugging Face, with APIs offered for developers.
Features
A versatile AI model with advanced Indic skills
Sarvam-M is a single, versatile model that supports both "think" and "non-think" modes.
The think mode is for complex logical reasoning, mathematical problems, and coding tasks, while the non-think mode is for efficient general-purpose conversation.
The model has been specifically post-trained on Indian languages with English, reflecting Indian cultural values authentically.
It also offers full support for Indic scripts as well as romanized versions of Indian languages.