GLOBAL RESEARCH ARCHIVE
India IT Services: Looking at strategic aspects of HCL’s investment into Sarvam AI
Research evidence excerpt
India IT Services: Looking at strategic aspects of HCL’s investment into Sarvam AI
A brief on Sarvam AI:
Sarvam is a private, unlisted company. The profile below is on basis of information on
company website and media articles
What does Sarvam AI do?
Sarvam AI focusses on building sovereign, full stack AI infrastructure for India. It is a
Bengaluru-based artificial intelligence company, founded in 2023 with offerings
spanning foundational models, inference infrastructure, and enterprise applications.
The company operates as a full-stack AI platform provider, offering tools for enterprises
and governments to build and deploy AI solutions across cloud, private cloud, and on-
premise environments.
It is positioned as an integrated model + platform + application provider, not too
dissimilar to the global peers building foundational AI ecosystems.
Snapshot of its product stack and technology
Sarvam has developed a multi-layer AI stack, anchored by:
1) Foundational Models (Core IP)
• Sarvam-105B (flagship): 105 billion parameters multilingual LLM for
enterprise-scale reasoning tasks.
• Sarvam-30B: optimized for edge and real-time use cases
Models trained from scratch on Indian datasets rather than fine-tuned Western models.
These models support 22 Indian languages, addressing a key capability gap vs. global
LLMs.
2) Speech, Vision, and Multimodal Systems
• Speech-to-text (e.g., Saaras V3) and text-to-speech (Bulbul V3)
• Vision-language model (Sarvam Vision) for document parsing and OCR
• Translation models for 22+ Indian languages
3) Platform + Application Layer
• Conversational AI agents (voice + chat)
• Document digitization workflows
• APIs and developer SDKs for enterprise integration
The architecture is designed for population-scale deployment, with “voice-first”
The English excerpt is extracted automatically from the cited source page and may contain layout or recognition errors. It is never batch translated.
Open report viewer