From ‘embarrassing’ to impressive: Top Silicon Valley VC calls Sarvam praises Indic AI breakthroughs

Deedy, a partner at Menlo Ventures, has significantly revised his earlier stance on Sarvam AI, an India-based startup specializing in artificial intelligence. In a recent post on X, he acknowledged the company's impressive advancements, describing them as "really valuable" and commending their technological execution. Initially skeptical about Sarvam's strategy to train smaller models for Indic languages, Deedy admitted that his perspective was misguided. He now recognizes the startup's robust progress in developing top-tier text-to-speech, speech-to-text, and optical character recognition (OCR) models tailored for Indic languages. He emphasized the startup's competitive pricing, user-friendliness, and refined product design, noting that Sarvam is effectively addressing a market gap often overlooked by larger global AI firms. Deedy clarified that he lacks detailed insights into Sarvam's business performance but highlighted the technical milestones the company has achieved. He expressed his astonishment, stating he could not recall the last time a software product from India made such a significant impact. This newfound admiration sharply contrasts Deedy's previous critiques, particularly from May 2025, when he labeled Sarvam’s initial flagship language model as underwhelming. At that time, he pointed out that the model, which featured 24 billion parameters and was trained on Indic data, had only managed 23 downloads within two days of its launch, comparing it unfavorably to a Korean model that garnered nearly 200,000 downloads in the same period. The tide began to turn with the introduction of Sarvam Vision, an advanced multimodal AI model released in February. This model, based on a 3-billion-parameter vision-language architecture, specializes in document intelligence, successfully integrating OCR, layout comprehension, and visual reasoning across India's multifaceted scripts and languages. In benchmark evaluations, Sarvam Vision achieved an impressive 84.3% accuracy on English documents and an average word accuracy of 87.36% across 22 Indian languages, surpassing global competitors like Google’s Gemini 3 Pro in Indic OCR tasks. The model excels in handling complex documents, including PDFs, scientific articles, and mixed-layout texts, areas where traditional OCR systems often falter. Founded in 2023 by experts from India's Aadhaar digital identity initiative, Sarvam AI is currently providing free access to its Document Intelligence APIs and Vision experience until February 2026, with the goal of fostering the growth of Indic-first AI applications.