A Bengaluru-based startup is challenging the Silicon Valley duopoly, proving that AI tailored for local context can outperform the world’s most capitalised models on their own turf
On 9 February 2026, the global AI community turned its eyes toward Bengaluru as Sarvam AI unveiled results showing its specialised models outperforming Google’s Gemini 3 Pro and OpenAI’s GPT-4o in key tasks. This breakthrough arrived just days before the India-AI Impact Summit 2026, where global tech leaders are gathering to discuss the democratisation of compute and intelligence. The startup’s latest vision model, Sarvam Vision, achieved a record 84.3 per cent accuracy on the olmOCR-Bench, a critical benchmark for document intelligence. This score significantly surpassed Gemini 3 Pro and DeepSeek OCR v2, while ChatGPT ranked notably lower.
Precision Over Parameters in Document Intelligence
By focusing on the unique messiness of Indian documents, which often feature complex layouts, mathematical formulas, and “code-mixed” text—Sarvam has effectively demonstrated the power of sovereign AI over general-purpose global models. Unlike the trillion-parameter “black box” models of Silicon Valley, Sarvam Vision is an inference-efficient, 3-billion-parameter state-space model. Its architecture enables high-fidelity knowledge extraction rather than mere text extraction. On the OmniDocBench v1.5, Sarvam Vision scored a staggering 93.28 per cent, excelling in parsing technical tables and handwritten notes that traditionally break foreign OCR systems.
Building the Infrastructure for 22 Official Languages
The startup has also released the Sarvam Indic OCR Bench, the first comprehensive evaluation framework for all 22 official Indian languages. By providing over 20,000 samples of varying quality—ranging from modern digital scans to 19th-century manuscripts—Sarvam is building the digital infrastructure necessary for India to digitise its vast administrative and historical archives. This level of accuracy was previously impossible under foreign-built systems that lacked the nuanced training data required for Indic scripts and various regional dialects.
Leading the Voice-First Revolution in Rural India
In a country where voice is the primary interface for millions, Sarvam’s Bulbul V3 has set a new standard for text-to-speech performance. Launched in early February 2026, Bulbul V3 outperformed global competitors in 8 kHz telephony-grade audio, which is the standard for Indian customer service and banking calls. The model handles the natural way Indians switch between languages without the robotic pauses or mispronunciations common in Western models. Furthermore, its low-latency streaming mode ensures responsiveness in conversational agents, making it ideal for rural sectors where connectivity can be intermittent.
Also Read: OpenAI Acquires Torch in $100M Deal to Empower ChatGPT Health
Practical Utility in National Policy Dissemination
The practical utility of this technology was recently showcased during the presentation of the Union Budget 2026. Finance Minister Nirmala Sitharaman’s speech was dubbed into Kannada and Hindi in real time using Sarvam Dub, achieving a 6.6-fold reduction in latency compared to standard implementations. This marked the first time a national budget was made accessible in real time across multiple regional languages via AI, ensuring that financial policy was transparent to non-English-speaking audiences. Supported by the government’s ₹10,300-crore IndiaAI Mission, Sarvam is training its models on domestic compute infrastructure, such as Yotta’s Shakti Cloud, to ensure data remains within national borders.
The Hinge Point
The 9 February 2026 performance of Sarvam AI is the exact moment where India moves from an “AI consumer” to an “AI creator.” This is the hinge point because it shows that a compact, 3B-parameter model trained on domain-specific data can defeat a general-purpose giant on regional tasks. The story changes here because it shatters the myth that only “hyperscale” compute and Silicon Valley labs can produce world-class intelligence. What can no longer remain the same is the dependency of Indian enterprises on foreign APIs for local language tasks. By achieving superior results at a fraction of the cost, Sarvam has made sovereign AI a business necessity. This marks the end of the “one model fits all” era and the beginning of the “local-first” AI revolution.