Sarvam-M: The Brain Behind India’s Multilingual AI 🧠

June 7, 2025 (1mo ago)

Fig: Cover Image

Introduction

In a world where artificial intelligence (AI) is transforming how we communicate, learn, and work, most language models are built with English or a handful of global languages in mind. This leaves millions of people, especially in linguistically diverse regions like India, underserved. Sarvam AI, a pioneering Indian startup focused on building a sovereign AI ecosystem for India, has taken a bold step to address this gap with Sarvam-M, a hybrid reasoning model designed specifically for Indic languages.

This blog post explores the development, performance, and potential impact of Sarvam-M, drawing insights from Sarvam AI’s technical blog. (Sarvam AI Blog)

🤖 What is Sarvam-M?

Sarvam-M is a large language model (LLM) designed to serve India’s diverse linguistic landscape, supporting 10 major Indian languages—Hindi, Bengali, Gujarati, Kannada, Malayalam, Marathi, Oriya, Punjabi, Tamil, and Telugu—which collectively cover about 70% of the country’s population. Built on the 24-billion parameter Mistral Small framework and licensed under Apache 2.0, Sarvam-M is tailored to excel not only in native scripts but also in code-mixed and romanized forms, making it highly adaptable to real-world communication patterns in India. In addition to its linguistic capabilities, the model is optimized for tasks such as coding, math, and culturally relevant reasoning, positioning it as a powerful and versatile tool for education, technical innovation, and diverse use cases across the country.

The model is accessible to the public in multiple ways:

🧪 The Development of Sarvam-M

Creating a model like Sarvam-M required a sophisticated blend of data curation, fine-tuning, and optimization. Here’s a breakdown of the key processes involved:

  1. Supervised Fine-Tuning (SFT)
  2. Reinforcement Learning with Verifiable Rewards (RLVR)
  3. Inference Optimizations

1. Supervised Fine-Tuning (SFT) 🎯

To make Sarvam-M proficient in Indic languages and specialized tasks, Sarvam AI undertook an extensive supervised fine-tuning process:

Fig: SFT Training Image

2. Reinforcement Learning with Verifiable Rewards (RLVR) ♻️

To further enhance Sarvam-M’s capabilities, Sarvam AI employed reinforcement learning with verifiable rewards (RLVR):

3. Inference Optimizations ⚙️

To ensure Sarvam-M performs efficiently in real-world applications, the team implemented several optimizations:

📊 Benchmarks Highlights

The benchmark comparison evaluates the performance of several large language models - Sarvam-M (24B), Mistral Small (24B), Gemma 3 (27B), Llama 4 scout (17B/109B), and Llama 3.3 (70B)—across diverse categories: General Knowledge, Programming, Math, Indic languages, and an overall “Vibe Check.” Sarvam-M emerges as the most well-rounded performer, particularly excelling in India-centric tasks such as MMLU-IN, ARC-C-IN, and MILU-IN, as well as dominating programming benchmarks like MBPP and LivecodeBench. Llama 3.3 leads global general knowledge tasks (MMLU), while Llama 4 Scout tops in math benchmarks like GSM-8 K. Gemma 3 holds strong in math and programming, especially on the MATH dataset. Sarvam-M also secures the highest Vibe Check score, indicating a balanced and user-aligned model experience. In contrast, Mistral Small lags behind most models across categories.

⚖️ Comparisons:

Fig: Benchmarks Image

🧱 Challenges and Lessons Learned

Developing Sarvam-M was not without its hurdles. One significant challenge was the low initial adoption rate, with only 334 downloads in the first two days on Hugging Face, compared to a Korean model’s 200,000 downloads (Analytics India Mag). Investor Deedy Das called this response “embarrassing,” arguing there was little audience for such incremental work, especially with more advanced models from Google and TWO.ai available.

Despite raising $41 million and being valued at $111 million (Sarvam AI Series A), some critics felt Sarvam AI’s contributions did not match the funding. Technically, while Sarvam-M excelled in Indian language and math tasks, it saw a 1% decline in English knowledge evaluations like MMLU, highlighting trade-offs in specializing in Indic languages.

The need for extensive Indic language data collection posed another challenge, as large-scale datasets are resource-intensive to curate. Community feedback was mixed, with some defending Sarvam-M’s methodology and others questioning the nationalism angle. Comparisons with other government-backed models, like BharatGen’s Param-1-1 with only 12 downloads (AIKosh), underscored similar adoption challenges.

Lessons learned include the importance of user engagement, clear communication of the model’s value, and the need for more extensive pre-training to balance Indic and English performance. These insights are crucial for future iterations and demonstrate Sarvam AI’s commitment to transparency.

🚀 The Future of AI in India

Sarvam-M is a cornerstone in building a sovereign AI ecosystem for India, aligning with Sarvam AI’s vision to make India an active participant in AI development, not just a consumer. Leveraging India’s strengths—government-led digital public goods like Aadhaar and UPI, a growing developer community, and ROI-focused enterprises—Sarvam AI is creating AI solutions tailored to India’s unique needs.

The company is developing a comprehensive Generative AI stack, including:

Supported by MeitY’s IndiaAI Mission and in collaboration with IIT Madras, Sarvam AI is fostering innovation through open-source initiatives and the Sarvam Circle, inviting global collaboration for research, custom models, technical education, and impact sector applications.

Sarvam-M’s potential applications are transformative:

As Sarvam AI continues to refine its models, the future of AI in India looks promising, with the potential to bridge digital divides, foster innovation, and drive economic growth.

Conclusion

Sarvam-M is more than a technical achievement; it’s a step toward making AI inclusive and relevant for India’s diverse population. Its open accessibility, robust performance, and focus on Indic languages position it as a catalyst for innovation. As Sarvam AI continues to refine and expand this model, it promises to reshape how AI serves one of the world’s most linguistically rich regions.

📚 Key Citations:

1. Sarvam AI Blog: Explorations in Post Training and Inferencing Optimizations
2. Sarvam-M Model on Hugging Face
3. Sarvam AI Playground for Testing Sarvam-M
4. Sarvam AI API Documentation
5. FAISS GitHub for Clustering Prompts
6. Alibaba gte-Qwen2-7B-instruct Model
7. Indic Evals Collection on Hugging Face
8. Open Sourcing R1-theory of the case

LinkedIn
X