Key highlights:

• Sarvam unveiled 30B and 105B parameter AI models at the India AI Impact Summit in New Delhi.
• The models use mixture-of-experts architecture to cut costs and support real-time use.
• Sarvam also announced speech, vision, and edge AI deployments for phones, cars, and smart glasses.
• The company plans to open source its largest models and focus on practical, India-first AI use cases.

Indian AI lab Sarvam has launched a new generation of artificial intelligence models and platforms at the India AI Impact Summit, marking its most ambitious push yet to build efficient, open source AI systems tailored for India.

The announcement includes two large language models, speech and vision systems, and plans to bring AI to feature phones, cars, and smart glasses.

Sarvam and India’s push for homegrown AI

The launch aligns closely with India’s broader effort to reduce reliance on foreign AI platforms and develop systems that work across Indian languages, accents, and local contexts.

Sarvam’s executives positioned the new models as an alternative to large, expensive AI systems developed in the U.S. and China. Instead of chasing size alone, the company is betting that smaller, more efficient models can power real-world applications at lower cost.

Founded in 2023, Sarvam has focused on voice-first AI and enterprise use cases. The latest announcements signal a broader shift toward consumer devices and everyday usage.

What models did Sarvam launch?

Sarvam unveiled a full AI stack rather than a single model. The lineup includes:

• A 30-billion-parameter large language model
• A 105-billion-parameter large language model
• A text-to-speech model
• A speech-to-text model
• A vision model designed to parse documents and images

These systems replace Sarvam’s earlier 2-billion-parameter Sarvam 1 model released in October 2024.

Why mixture-of-experts matters for Sarvam

Both the 30B and 105B models use a mixture-of-experts architecture. This design activates only a small portion of the model’s total parameters during each task.

Sarvam said this approach significantly reduces computing costs while maintaining strong performance. It also makes the models more practical for real-time applications.

The 30B model supports a 32,000-token context window. Sarvam said it is optimized for conversational and voice-based systems.

The 105B model offers a 128,000-token context window. This allows it to handle longer documents and complex, multi-step reasoning tasks.

How Sarvam’s models compare globally

Sarvam directly positioned its models against global peers.

The 30B model is placed alongside:

• Google Gemma 27B
• OpenAI GPT-OSS-20B

The larger 105B model is aimed at competing with:

• OpenAI’s GPT-OSS-120B
• Alibaba Qwen-3-Next-80B

Sarvam emphasized that its models were trained from scratch, not fine-tuned on existing open source systems.

How were the models trained?

Sarvam said the 30B model was pre-trained on about 16 trillion tokens of text. The 105B model was trained on trillions of tokens spanning multiple Indian languages.

The training used computing resources provided under India’s government-backed IndiaAI Mission.

Infrastructure support came from data center operator Yotta, with technical support from Nvidia. Sarvam did not disclose detailed training data sources or full compute costs.

Will Sarvam open source its models?

Sarvam said it plans to open source both the 30B and 105B models. However, the company did not clarify whether it will release the full training datasets or complete training code.

Open sourcing the models could allow developers, startups, and researchers to build local applications without depending entirely on foreign AI platforms. The move also fits India’s broader narrative around sovereign AI and data control.

What real-world use cases is it targeting?

Sarvam said the new models are designed for real-time applications. These include:

• Voice-based assistants
• Chat systems in Indian languages
• Customer support automation
• Document processing

Company executives said they plan to scale carefully and focus on tasks that matter at production scale.

“We want to be mindful in how we do the scaling,” co-founder Pratyush Kumar said at the event. “We don’t want to do the scaling mindlessly.”

Sarvam for Work and Samvaad explained

Alongside the core models, Sarvam outlined two product directions. Sarvam for Work is positioned as an enterprise-focused suite. It will include coding-focused models and tools designed for business workflows. Samvaad is Sarvam’s conversational AI agent platform. It aims to help developers build multilingual assistants that can work across voice and text.

The company did not share timelines for broader availability.

Bringing AI to feature phones and offline devices

One of the most notable announcements was Sarvam’s push into edge AI. The company said it is working with HMD to bring a conversational AI assistant to Nokia and HMD feature phones. A demo showed a dedicated AI button that lets users speak in a local language.

The assistant can offer guidance on government schemes and local market information. Sarvam did not confirm which features will work fully offline. Sarvam said its edge models are small enough to run on devices with limited processing power and storage.

Partnerships with Qualcomm and Bosch

Sarvam also announced collaborations to expand its AI footprint beyond phones.

The company said it has worked with Qualcomm to tune its models for Qualcomm chipsets. Qualcomm is developing a Sovereign AI Experience Suite for phones, PCs, cars, and IoT devices.

Sarvam is also working with Bosch to bring AI assistants to cars. Details about deployment timelines were not shared.

These partnerships suggest Sarvam wants its models to live closer to users, not just in the cloud.

Kaze smart glasses: What we know

Sarvam also showcased a pair of AI smart glasses called Sarvam Kaze. The company said the device is designed and manufactured in India.

Co-founder Pratyush Kumar described it as a “builders’ device.” Sarvam said the glasses will be available in May. The company did not share pricing or detailed specifications at the event.

Funding and investors

Sarvam has raised more than $50 million since its founding in 2023.

Its investors include:

• Lightspeed Venture Partners
• Khosla Ventures
• Peak XV Partners

Until now, the company has largely focused on enterprise deployments, especially voice AI for customer support. The new announcements suggest a broader consumer push.

What comes next?

Sarvam said it will continue to take a measured approach to scaling. The company plans to refine its models based on real-world feedback and expand partnerships across devices and industries.

As India pushes for sovereign and locally relevant AI, Sarvam’s roadmap places it at the center of that conversation.

In the coming months, the real test will be how widely these models are adopted and how well they perform outside controlled demos.

Conclusion

At the India AI Impact Summit, Sarvam laid out its most comprehensive AI strategy yet, spanning large language models, speech systems, and edge deployments. By betting on efficient, open source AI built for Indian realities, Sarvam is signaling how it plans to compete in a global market dominated by much larger players.