Features
On-device inference with sub-120ms latency
Hybrid cloud routing based on model confidence
Automatic cloud fallback for complex/noisy requests
Transcription with <6% WER and privacy mode
Tool calling and function calling (Needle 26M model)
Voice activity detection (Silero VAD)
Multi-platform SDK (iOS, Android, macOS, wearables)
INT4/INT8 quantization with zero-copy memory mapping
NPU acceleration on Apple, Snapdragon, Exynos, MediaTek
OpenAI-compatible API endpoints
Cactus Graph for custom model implementation
Cactus Kernels: custom attention with KV-cache quantization
TurboQuant-H: 2-bit embedding quantization for Gemma 4
Needle 26M distilled model for high-speed tool calling
Offline-capable inference mode
Text to Speech in 11 Indic languages with natural voices
Speech to Text (ASR) across 12 Indic languages
Translation across 23 languages
Document digitisation from PDFs and images into structured data
Dubbing capabilities for multimedia
Conversational AI agents (Samvaad) for voice and chat
Akshar document digitisation product
Arya AI for Work platform for enterprise workflows
Sovereign deployment: cloud, VPC, on-premises, air-gapped
Forward deployed engineering support for custom solutions
SOC 2 Type II, ISO 27001, DPDP Act compliant
Role-based access control and full audit trails
Models: Bulbul (TTS), Saaras (ASR), Mayura (Doc), Indus (LLM)
Python SDK (pip install sarvamai) and REST API
Playground for browser-based testing without setup
