In a breakthrough for “on-device” intelligence, Bengaluru-based Sarvam AI officially launched Sarvam Edge on February 14, 2026. This innovative AI stack allows advanced generative AI tasks—such as speech recognition, translation, and document parsing—to run entirely offline on consumer hardware like smartphones and laptops.
By shifting inference from the cloud to the device, Sarvam Edge addresses three of India’s biggest tech hurdles: high cloud costs, inconsistent rural connectivity, and data privacy concerns.
The Technical Edge: Efficiency Over Scale
Sarvam Edge is designed to be “super small” in footprint without sacrificing the accuracy typically associated with massive server-side models.
| Component | Parameter Count | Footprint (FP16) | Key Capability |
| Speech-to-Text | 74 Million | ~294 MB | Transcription accuracy rivaling cloud systems. |
| TTS (Bulbul Edge) | Unified Model | Compact | Intelligible speech in 10 Indian languages. |
| Translation | Unified Model | Efficient | Bidirectional across 110 language pairs. |
| Vision (OCR) | 3 Billion | ~10 GB RAM | 40+ tokens/sec on consumer hardware. |
Key Features of the Edge Stack:
- Instant Response: No round-trips to data centers means zero latency and no “queuing” behind other users.
- Sovereign Privacy: Since the data never leaves the device, there is no risk of server-side logging or data leaks.
- Multilingual Native: Supports 11 languages (10 Indian + English) out of the box, with built-in language identification and code-switching (Hinglish, etc.) support.
- Zero Variable Cost: By removing per-query API costs, developers can offer AI features to millions of users without incurring exponential cloud bills.
Real-World Performance: “Sarvam Vision” in Action
During a live demonstration on February 15, the company showcased Sarvam Vision OCR running on a standard MacBook Pro with the internet completely disabled. The model transcribed a complex Odia document image at speeds exceeding 40 tokens per second, proving that local hardware is now capable of handling specialized Indian language tasks that previously required massive GPU clusters.
Strategic Context: Part of the $2.3B Sovereign Vision
The launch of Sarvam Edge follows the massive $2.3 billion MoU signed with the Odisha Government to build a “Sovereign AI Capacity Hub.” While the Odisha hub provides the national compute backbone, Sarvam Edge serves as the local execution layer, bringing that intelligence to the hands of citizens in rural and tribal areas where internet bandwidth is often limited.
“Intelligence should work everywhere. Not summoned from distant servers, not gated behind connectivity, not metered by the query. Just there, immediate and local.” — Sarvam AI Blog.
