In a landmark development for India’s artificial intelligence landscape, Bengaluru-based startup Sarvam AI has unveiled its flagship Large Language Model (LLM) named Sarvam-M. This 24-billion-parameter hybrid language model is designed to excel in multilingual understanding, mathematical reasoning, programming tasks, and Indian languages, setting new benchmarks in these domains for a model of its size.
What is Sarvam-M?
Sarvam-M is an open-weights, hybrid-reasoning, text-only language model built on top of Mistral Small, an efficient open-weight AI model developed by the French AI firm Mistral AI. The model is open source and accessible to developers via Hugging Face, with APIs provided for easy integration into various applications.
The “M” in Sarvam-M stands for Mistral, reflecting its foundational architecture. Sarvam AI developed this model as part of the Indian government’s IndiaAI Mission, aiming to build a sovereign AI ecosystem in India and make generative AI accessible at scale.
Key Features and Innovations
- Hybrid Thinking Mode: Sarvam-M supports two operational modes —
- Think mode for complex logical reasoning, math problems, and programming tasks.
- Non-think mode for general-purpose conversations and smooth chat interactions.
This dual-mode design allows the model to handle a broad spectrum of use cases, from casual dialogue to sophisticated problem-solving.
- Advanced Indian Language Support: The model is specifically post-trained on a rich dataset of Indian languages alongside English. It supports multiple Indic scripts as well as romanized versions of Indian languages, reflecting Indian cultural values authentically and enhancing accessibility for diverse linguistic users.
- Superior Reasoning and Performance: Sarvam-M has demonstrated exceptional performance improvements over its base model and many contemporaries:
- 20% average improvement on Indian language benchmarks.
- 21.6% enhancement on math-related tasks.
- 17.6% improvement in programming benchmarks.
- An impressive +86% gain in romanized Indian language math benchmarks (GSM-8K).
- Competitive Edge: Sarvam-M outperforms Meta’s LLaMA-4 Scout on most benchmarks and is comparable to larger models like LLaMA-3.3 70B and Google’s Gemma 3 27B, despite being trained on fewer tokens. However, it slightly lags (~1% point) on English knowledge benchmarks such as MMLU.
Development Process
Sarvam AI employed a rigorous three-step training and optimization pipeline to build Sarvam-M:
- Supervised Fine-Tuning (SFT):
The team curated a wide range of high-quality, challenging prompts. They generated completions using permissible models, filtered these outputs through custom scoring mechanisms, and adjusted them to reduce bias while ensuring cultural relevance. This process trained the model to handle both complex reasoning (“think”) and general conversation (“non-think”) modes effectively. - Reinforcement Learning with Verifiable Rewards (RLVR):
Sarvam-M underwent further training on a curriculum combining instruction-following, programming datasets, and mathematical problems. Custom reward engineering and prompt sampling strategies were used to enhance task-specific performance. The training algorithm employed was GRPO, optimized with hyper-parameter tuning. - Inference Optimizations:
Post-training quantization to FP8 precision was applied, achieving negligible accuracy loss while improving efficiency. Techniques like lookahead decoding were implemented to boost throughput, although challenges remain in supporting higher concurrency during inference1.
Applications and Use Cases
Sarvam-M is designed for versatility, targeting a broad range of applications including:
- Conversational AI: Natural, context-aware chatbots capable of handling Indian languages fluently.
- Machine Translation: Accurate translation services across Indic languages and English.
- Educational Tools: Assisting students and learners with math, programming, and language tasks.
- Programming Assistance: Code generation, debugging, and explanation in multiple programming languages.
- Reasoning Tasks: Solving complex mathematical problems and logical reasoning challenges.
Significance for India and the AI Ecosystem
Sarvam AI’s launch of Sarvam-M marks a significant milestone in India’s journey toward AI sovereignty. By open-sourcing such a powerful LLM tailored to Indian languages and cultural context, Sarvam AI is fostering an ecosystem that empowers developers, researchers, and enterprises within India to build localized AI solutions without heavy reliance on foreign models.
The company has committed to regular model updates and sharing detailed technical findings, signaling ongoing innovation and community engagement in India’s AI space.
Conclusion
Sarvam-M stands out as a pioneering open-source LLM that combines scale, versatility, and deep Indian language expertise. It not only competes with global AI giants but also addresses the unique linguistic and cultural needs of India. As Sarvam AI continues to refine and expand its capabilities, Sarvam-M is poised to become a foundational technology for India’s AI-driven future.