Sarvam AI launches Sarvam-M, India's open-weights model for sovereign AI

Sarvam AI, a prominent Indian AI startup, has introduced Sarvam-M, a 24-billion parameter open-weights model, marking a significant development in the country's pursuit of sovereign AI capabilities. This new model is built on top of Mistral Small and demonstrates notable advancements in handling Indian languages, mathematics, and programming tasks.[1] The launch positions Sarvam-M as a powerful tool for a wide array of applications, including conversational AI, machine translation, and educational tools, and it is accessible via the Sarvam API.[1] This initiative is part of Sarvam AI's broader strategy to foster a robust AI ecosystem within India, with further announcements anticipated.[1]

Sarvam AI has been actively working towards creating AI solutions tailored to India's diverse linguistic landscape and is recognized as a key player in the IndiaAI Mission, a government initiative to build the nation's foundational Large Language Model (LLM).[1][2] The company's mission emphasizes making generative AI accessible and beneficial across India, aiming to prevent the nation from becoming merely a consumer of AI technologies developed elsewhere.[3][4] Co-founder Vivek Raghavan has stated that Sarvam-M represents an important step in their journey to build Sovereign AI for India.[1] The company has previously released other models, including OpenHathi, a Hindi LLM, and Bulbul, a speech AI model supporting 11 Indian languages with natural, region-specific accents.[1][4] Sarvam AI's strategy involves both creating LLMs from scratch and leveraging existing open-source models.[4] They are also collaborating with AI4Bharat at IIT Madras for building various model variants.[2] The long-term vision includes developing models at multiple scales and modalities, such as Sarvam-Large for complex tasks, Sarvam-Small for real-time applications, and Sarvam-Edge for on-device processing.[3][2]

Sarvam-M itself is a multilingual, hybrid-reasoning, text-only language model that shows substantial improvements over the base Mistral Small model it was built upon.[5] According to Sarvam AI, these improvements include a 20% average enhancement on Indian language benchmarks, a 21.6% increase on math benchmarks, and a 17.6% boost on programming benchmarks.[5][6] Notably, the performance gains are even more significant at the intersection of Indian languages and mathematics, with an 86% improvement in romanized Indian language GSM-8K benchmarks.[5][6] To achieve this, approximately one-third of the training samples were generated in Indic languages.[1] Specifically, 30% of coding, math, and reasoning prompts, and 50% of other prompts were translated into Indian languages.[1][6] The model focuses on ten key Indic languages: Hindi (which accounted for 28% of the Indic data), Bengali, Gujarati, Kannada, Malayalam, Marathi, Oriya, Punjabi, Tamil, and Telugu, with each of the latter nine languages constituting 8% of the Indic data.[1][6] These ten languages collectively represent the mother tongue of over 70% of India's population, roughly one billion people.[1][6] Sarvam-M features a "Hybrid Thinking Mode," allowing it to switch between complex logical reasoning for tasks like math and coding, and a more efficient mode for general conversation.[5] It also offers full support for both Indic scripts and romanized versions of Indian languages.[5] The company detailed its methodology, which includes supervised fine-tuning (SFT), reinforcement learning with verifiable rewards (RLVR), and inference optimizations, in a technical blog accompanying the launch.[1][6]

The emphasis on coding and mathematical reasoning is a core strength of Sarvam-M.[1] It reportedly outperforms most similarly-sized models on these benchmarks.[5] For instance, Sarvam-M has shown high scores on programming benchmarks like HumanEval (0.88), MBPP (0.75), and LivecodeBench (0.44).[6] In mathematical reasoning, it achieved top performance on GSM-8K-IN (0.92) and GSM-8K-IN-R (0.82), while maintaining competitive scores on GSM-8K (0.94) and the MATH benchmark (0.81).[6] This strong performance in coding and reasoning, combined with its Indic language capabilities, makes Sarvam-M a potentially transformative tool for Indian developers and enterprises.[5] The model's ability to understand and generate code, alongside its proficiency in local languages, could lower entry barriers for non-English speaking programmers and foster more localized software development. Sarvam AI has also developed "Indic Vibe Check" to better evaluate how users would engage with the model in culturally relevant contexts.[6]

The launch of Sarvam-M has broader implications for the AI industry, particularly in India. It signifies a move towards greater self-reliance in critical AI technology and aligns with the Indian government's push for "Sovereign AI".[1][3] By providing an open-weights model proficient in regional languages and technical tasks, Sarvam AI is contributing to the democratization of AI tools and potentially catalyzing innovation across various sectors in India, from education to enterprise solutions.[1][7] The availability of such models can empower local developers and businesses to create AI applications tailored to unique Indian needs without concerns about data sovereignty or dependence on foreign technologies.[3][2] Furthermore, Sarvam AI's commitment to regularly releasing models and sharing technical findings is likely to stimulate further research and development within the Indian AI community.[6] This development comes at a time when India is significantly ramping up its AI ecosystem, including government initiatives to provide substantial GPU resources.[3][8] Sarvam AI's efforts, alongside those of other domestic players and partnerships with global tech companies like Microsoft, are crucial in shaping India's trajectory as not just a consumer, but also a co-creator of advanced AI technologies.[7][3][4][9]

In conclusion, Sarvam AI's release of Sarvam-M is a noteworthy advancement in the field of generative AI, especially for India. By building upon a strong open-source foundation like Mistral Small and significantly enhancing it for Indic languages, coding, and mathematical reasoning, Sarvam AI has delivered a versatile and powerful tool.[1][5] This open-weights model is poised to accelerate the development of AI applications tailored to the Indian context, foster innovation within the local developer community, and contribute significantly to India's ambition of achieving self-sufficiency and leadership in artificial intelligence.[1][3][6] The focus on regional languages and robust technical capabilities addresses key needs in the Indian market and sets the stage for a more inclusive and empowered AI-driven future for the nation.