NVIDIA 共同設計大幅提升 Sarvam 推論

• Mixture of Experts (MoE) Architecture: Sarvam 30B model activates only 1B of 30B parameters per output token; 105B model activates 9B parameters, reducing computational overhead and inference latency[2] • Training Scale: 30B model trained on 16 trillion tokens with 32,000-token context window; 105B model trained on 17+ trillion tokens with 128,000-token context window[2] • Hardware Foundation: Deployed on NVIDIA H100 SXM GPUs (4,096 units allocated to Sarvam)[2] • Specialized Models: Sarvam Vision (3B parameters) for document intelligence and OCR; Saaras V3 for Indic speech recognition achieving 19.3% word error rate on IndicVoices benchmark covering ten major Indian languages[5] • Co-design Integration: NVIDIA's hardware-software co-design optimizes inference for conversational and voice-based AI agents requiring high throughput and predictable latency[1] • Production Infrastructure: Pravah platform enables production-grade inference for government and enterprise applications[4]

Sarvam AI's efficient model architecture and NVIDIA co-design partnership position India's sovereign AI capabilities as competitive alternatives to global frontier models, particularly for multilingual and document-intensive workloads. The success of government-subsidized foundational model development through IndiaAI Mission (expanded from 4 to 12 startups by February 2026) demonstrates viability of domestic AI infrastructure independent of foreign systems. The 128,000-token context window and superior performance on Indian language tasks suggest emerging market differentiation in regional AI services. However, Sarvam's acknowledged limitations outside specialized domains (OCR, speech, document intelligence) indicate the company must validate its upcoming 120B sovereign model as a true general-purpose competitor to GPT, Gemini, and Claude to justify its positioning as a comprehensive alternative to global AI leaders.