Indian AI Lab Sarvam Launches Open-Source Models to Challenge U.S. and Chinese Rivals

Indian artificial intelligence startup Sarvam on Tuesday released a suite of large language models designed to compete with systems from OpenAI and Alibaba, marking a significant escalation in the country's efforts to build homegrown AI infrastructure that doesn't depend on foreign platforms.

The launch, announced at the India AI Impact Summit in New Delhi, represents a major technical leap for the Bangalore-based company and a test of whether smaller, efficient open-source models can capture market share from the expensive proprietary systems dominating corporate AI deployments. For finance leaders evaluating AI vendors, the release signals a potential shift in the competitive landscape—and a new set of questions about whether "good enough and cheap" can beat "best in class and expensive."

Sarvam's new lineup includes two flagship models—a 30-billion-parameter system and a 105-billion-parameter model—alongside specialized tools for text-to-speech, speech-to-text, and document parsing. The models represent a sharp upgrade from the company's 2-billion-parameter Sarvam 1 model released in October 2024, and were trained from scratch rather than fine-tuned on existing open-source systems.

The technical architecture is where things get interesting for CFOs thinking about AI economics. Both flagship models use a mixture-of-experts design, which activates only a fraction of their total parameters at any given time. This approach, Sarvam said, significantly reduces computing costs—the kind of efficiency claim that matters when you're trying to justify AI spending to a board that's heard too many promises about "transformative" technology.

The 30-billion-parameter model was pre-trained on about 16 trillion tokens of text and supports a 32,000-token context window aimed at real-time conversational applications. The larger 105-billion-parameter model, trained on trillions of tokens spanning multiple Indian languages, offers a 128,000-token context window designed for more complex, multi-step reasoning tasks. Sarvam positions the 30B model against Google's Gemma 27B and OpenAI's GPT-OSS-20B, while the 105B model is touted to compete against OpenAI's GPT-OSS-120B and Alibaba's Qwen-3-Next-80B.

The launch aligns with New Delhi's broader push to reduce reliance on foreign AI platforms and develop models tailored to local languages and use cases. Sarvam said the models were trained using computing resources provided under India's government-backed IndiaAI Mission, with infrastructure support from data center operator Yotta and technical support from Nvidia. (Translation: this isn't just a startup story—it's industrial policy with silicon backing.)

For multinational finance teams, the strategic question is whether regional AI models become a compliance requirement or a competitive advantage. If India's government starts favoring locally-trained models for sensitive applications—think payroll, tax compliance, or regulatory reporting—CFOs with operations in the country may need to evaluate these systems regardless of whether they're "best in class" by Silicon Valley standards.

The models are designed to support real-time applications including voice-based assistants and chat systems in Indian languages, Sarvam said. That focus on multilingual capability and regional optimization represents a different value proposition than the English-first systems that dominate enterprise AI today.

The broader implication: the AI vendor landscape is fragmenting along both technical and geopolitical lines. Finance leaders who assumed they'd pick between OpenAI, Google, and Anthropic may find themselves managing a more complex portfolio of regional models, open-source alternatives, and specialized systems—each with different cost structures, compliance implications, and performance tradeoffs. The "one AI platform to rule them all" procurement strategy is looking increasingly naive.