Sarvam AI has announced two large language models—Sarvam 30B and Sarvam 105B—built from scratch using a Mixture of Experts (MoE) architecture. The release signals a clear push toward enterprise-grade AI while keeping efficiency and scale in focus.
What’s new in Sarvam 30B and 105B
Both models rely on MoE, a design that routes each request through a subset of specialized “experts” instead of activating the entire network. This approach delivers stronger performance per token while keeping compute costs under control.
The Sarvam 105B model activates 9B parameters per token and supports a 128K context window. That combination enables long-document reasoning, multi-step planning, tool use, and complex coding workflows without constant context trimming.
Built for demanding, real-world workloads
Sarvam positions the 105B model for:
- Complex reasoning and agentic task completion
- Coding, mathematics, and science workflows
- Tool-augmented applications that require long memory
- Population-scale and enterprise deployments
The 30B model targets lighter but still capable workloads where latency and cost matter, offering a practical option for production use.
Sarvam AI Confirms Open-Weight Models for Developers
Sarvam confirmed that both models will be released as open weights on Hugging Face, with API access and dashboard support to follow. Open weights lower the barrier for research, customization, and private deployments—an important move for teams that need control over data and infrastructure.
A full India-first AI platform
Sarvam’s broader platform covers:
- Speech-to-text with diarization and translation
- Text translation across 22 Indian languages plus English
- Natural text-to-speech voices
- Document intelligence with OCR and structured output
The new 30B and 105B models slot into this stack as the reasoning layer, enabling end-to-end AI systems designed for India’s linguistic diversity and real-world usage.
This release goes beyond a headline model size. Sarvam combines efficient MoE design, long-context reasoning, open weights, and predictable pricing into a single, production-ready offering. That mix makes the platform appealing to startups, enterprises, and public-sector teams that need scalable AI without opaque costs or closed systems.
What to watch next
- Availability of open weights on Hugging Face
- API limits and latency benchmarks
- Real-world comparisons with other long-context models
- Adoption by enterprises building India-first AI products
Sarvam’s 30B and 105B launch marks a significant step toward a homegrown, enterprise-ready AI ecosystem—one that emphasizes efficiency, openness, and practical deployment over raw parameter counts alone.
