India’s tech scene is buzzing with the launch of Sarvam-M, a 24-billion-parameter large language model (LLM) from Bengaluru-based startup Sarvam AI. Unveiled on May 23, 2025, this open-weights hybrid model, built on the French Mistral Small framework, is tailored for Indian languages, math, and programming tasks. Selected under the IndiaAI Mission to lead India’s sovereign AI efforts, Sarvam-M promises to bridge linguistic and digital divides. But with only 334 downloads in its first two days on Hugging Face, it’s also stirred controversy. Why is everyone talking about it? Let’s dive into what makes Sarvam-M a game-changer—and why it’s facing scrutiny.
What Is Sarvam-M? A Made-in-India AI Marvel
- Powerful Foundation: A 24-billion-parameter LLM built on Mistral Small, fine-tuned with supervised fine-tuning (SFT) and reinforcement learning with verifiable rewards (RLVR).
- Indian Language Focus: Supports 10 Indian languages, including Hindi, Bengali, Tamil, and Gujarati, with seamless handling of code-mixed and romanized text.
- Versatile Applications: Powers chatbots, translation tools, and educational platforms, excelling in math and programming tasks.
Sarvam-M is designed to resonate with India’s 1.4 billion people and 22 official languages. Unlike Western models like ChatGPT, which prioritize English, Sarvam-M is trained on Indian data to understand cultural nuances and colloquialisms. Its “Think” mode, for instance, aced JEE Advanced-level questions in Hindi, making it a potential game-changer for education. With a 20% improvement in Indian language benchmarks, 21.6% in math, and 17.6% in programming, it outperforms Meta’s Llama-4 Scout and rivals larger models like Llama-3.3 70B.
Why It Matters: A Step Toward Sovereign AI
- IndiaAI Mission Backing: Chosen in April 2025 to build India’s first sovereign LLM, Sarvam-M uses local compute resources for secure, scalable deployment.
- Breaking Language Barriers: Enables AI access for non-English speakers, from rural farmers to urban students, via voice-enabled and multilingual tools.
- Cost-Effective Innovation: Built efficiently to compete with global giants, aligning with India’s vision of tech independence by 2040.
Sarvam-M isn’t just another AI model—it’s a bold statement of India’s ambition to lead in AI. Co-founders Vivek Raghavan and Pratyush Kumar, veterans of AI4Bharat and Aadhaar, aim to create an AI ecosystem rooted in India’s linguistic diversity. From powering customer service for companies like Zepto to aiding government services, Sarvam-M’s applications could transform sectors like education, healthcare, and finance. Its open-source availability on Hugging Face and APIs encourages developers to innovate locally.
The Controversy: Why the Mixed Reception?
- Low Initial Downloads: Only 334 downloads in two days on Hugging Face, compared to 200,000 for Korea’s Dia model, sparked criticism.
- Expectations vs. Reality: Critics like Deedy Das called it “embarrassing,” arguing it lacks groundbreaking impact compared to global models like DeepSeek.
- Not Fully Homegrown: Built on Mistral Small, some question its “sovereign” label, preferring a fully scratch-built model.
The tepid response has fueled debates about India’s AI strategy. Critics argue Sarvam-M focuses too much on technical metrics rather than practical, India-specific problems like aiding farmers or local governments. Others, like AI4Bharat’s team, defend its methodology, emphasizing the rigorous training process over download numbers. Supporters see it as a stepping stone, with Sarvam’s upcoming 70-billion-parameter model promising a fully indigenous approach.
Real-World Impact: Where Sarvam-M Shines
- Education Boost: Solves complex math and science problems, supporting students in regional languages for exams like JEE.
- Business Efficiency: Voice-enabled AI agents for customer service in BFSI, healthcare, and retail, costing as low as ₹1 per minute.
- Cultural Relevance: Understands India’s linguistic diversity, making AI accessible to non-English speakers.
Sarvam-M’s ability to handle code-mixed Hinglish and regional dialects sets it apart. For instance, it translated a Lex Fridman podcast with PM Narendra Modi into nine Indian languages, broadening access. Its low-latency design and focus on voice-first interfaces make it ideal for India’s multilingual, mobile-first market.
Challenges Ahead: Can Sarvam-M Win Hearts?
- Adoption Hurdles: Low downloads suggest a need for better marketing and clearer use cases to excite developers and businesses.
- English Bias: Many Indians prefer English on devices, reducing demand for Indic-focused models.
- Competition: Faces pressure from global giants like OpenAI and cheaper models like Google’s Gemma, requiring unique value to stand out.
Critics argue Sarvam needs to tell a compelling story, like enabling elderly Indians to use tech or bringing regional poetry to life. Without emotional resonance, it risks fading like past India-first ventures such as Koo. Yet, supporters like Tarun Bhojwani argue Sarvam’s focus on execution over hype will pay off long-term.
The Road Forward: India’s AI Future
- Bigger Ambitions: Sarvam is developing a 70-billion-parameter multimodal model under the IndiaAI Mission, promising deeper innovation.
- Ecosystem Building: Partnerships with Nvidia, Microsoft Azure, and Yotta ensure robust infrastructure for scaling.
- Community Support: Open-sourcing encourages developers to build India-centric solutions, fostering a sovereign AI ecosystem.
Sarvam-M is a bold first step toward India’s AI sovereignty. While it faces skepticism, its focus on Indian languages and practical applications positions it as a catalyst for inclusive tech. As Pratyush Kumar said, “We’re building AI that reaches every corner of India.” With ongoing innovations like the Bulbul-v2 voice model and enterprise-focused Sarvam Agents, Sarvam AI is paving the way for a future where AI speaks India’s languages and solves its unique challenges.
