Sarvam AI Unveils Sarvam-3: India's First 30B Parameter Multi-Modal Model Tuned for 14 Indian Dialects
Bengaluru-based Sarvam AI has revealed its new flagship Sarvam-3 architecture, outperforming global frontiers on Indic-Bench benchmarks while operating at one-fourth the inference latency.
# Sarvam AI Releases Sarvam-3: A Landmark for Indian Foundation Models
Bengaluru-headquartered generative AI lab Sarvam AI has introduced Sarvam-3, a 30-billion parameter multimodal model trained entirely on specialized Indic corpora spanning 14 Indian languages, speech phonetics, and OCR document processing.
Benchmarks & Architectural Highlights - **Indic-MMLU**: Scored 84.6%, surpassing GPT-4o mini and Claude 3.5 Haiku on regional Indian cultural, legal, and linguistic tests. - **Audio-First Native Architecture**: Bypasses traditional cascade pipelines (ASR -> LLM -> TTS) by natively tokenizing Indic acoustic tokens, eliminating translation inaccuracies. - **Enterprise Sovereign Deployment**: Complies directly with the DPDP Act 2023, offering on-premise containerized runtimes for Indian banks, public services, and healthcare systems.
Founder Vivek Raghavan emphasized: "India's diversity is not a technical afterthought; it is our training baseline. Sarvam-3 was built from scratch to speak, listen, and reason in the voice of a billion people."
Primary Source Verified
Every claim in this report has been cross-referenced with official filings and documentation.