Sarvam AI becomes India's newest unicorn with $234m funding

🔑 Enhanced Key Takeaways

•Sarvam AI was founded in August 2023 by Vivek Raghavan and Pratyush Kumar, both veterans of AI4Bharat at IIT Madras, with Raghavan having prior experience in India's digital public infrastructure, including Aadhaar.
•Prior to its Series B, Sarvam AI had secured approximately $41 million in a combined seed and Series A funding round in December 2023, led by Lightspeed Venture Partners, with participation from Peak XV Partners and Khosla Ventures.
•The company has released open-source foundational models, including Sarvam-30B and Sarvam-105B (Mixture-of-Experts architecture), which were trained from scratch on datasets focused on Indian languages, and also offers multimodal systems like speech-to-text, text-to-speech (Bulbul V3), and vision-language models (Sarvam Vision).
•Sarvam AI was selected in April 2025 by the Ministry of Electronics and Information Technology (MeitY) to develop an indigenous foundational model under the IndiaAI Mission and has collaborated with the Unique Identification Authority of India (UIDAI) to integrate AI-based voice interactions into Aadhaar services.
•HCLTech's strategic investment of $150 million in the Series B round gives it a 10.46% stake in Sarvam AI, aiming to combine Sarvam's AI research with HCLTech's global enterprise presence to create a differentiated full-stack AI platform for enterprises and governments.

📊 Competitor Analysis▸ Show

Company/Platform	Focus/Key Features	Comparative Benchmarks/Notes
Sarvam AI	Sovereign, full-stack AI for India; LLMs, speech, vision, translation in 22+ Indian languages; enterprise & government solutions.	Sarvam Translate shows strong performance against larger models. Sarvam Vision claims leading performance on Indic OCR benchmarks against Gemini 3 Pro, Claude Opus 4.5, and GPT-5.2.
Krutrim AI	Domestic unicorn (2024), cloud-to-model ecosystem, consumer integration; focuses on NLP for local languages.	Krutrim Pro (150B parameters) processes 3 million tokens/second, outperformed by DeepSeek R-1.
Hanooman AI	Multimodal AI (text, speech, vision) in multiple Indian languages; led by IIT Bombay, supported by Reliance Jio.	Aims to make AI more accessible to Indian businesses.
CoRover's BharatGPT	Multilingual virtual assistant for Indian government services, banking, and e-commerce.	Supports regional Indian languages for chatbots and AI-driven customer support.
Google (Project Vaani)	Hyperscaler offering Indic language tools, bundled into cloud suites.	Competes on cloud distribution and integrated stacks, pressuring Sarvam's margins.
Microsoft (Bhashini)	Hyperscaler offering Indic language tools, bundled into cloud suites.	Competes on cloud distribution and integrated stacks, pressuring Sarvam's margins.
Global LLMs (ChatGPT, Gemini, DeepSeek, Claude)	General-purpose, global audience LLMs.	Sarvam AI is purpose-built for India's linguistic and operational realities, focusing on Indian-language performance where global models may incur a "tokenization tax" due to poor processing of Indian languages. DeepSeek R-1 (500B parameters) processes 10 million tokens/second.

🛠️ Technical Deep Dive

Foundational Models: Sarvam-30B and Sarvam-105B, both utilizing a Mixture-of-Experts (MoE) architecture.
Sarvam-30B: Features 30 billion parameters with approximately 1 billion active parameters per token generation. It employs a 19-layer depth (1 dense + 18 MoE) with 128 experts and a top-6 routing strategy, leveraging grouped query attention (GQA).
Sarvam-105B: Scales to 105 billion parameters with 10.3 billion active parameters. It uses a 32-layer depth (1 dense + 31 MoE) and employs top-8 routing over 128 experts with a larger MoE FFN hidden size of 2048. This model also adopts multi-head latent attention (MLA) for aggressive Key-Value (KV) cache compression.
Training Data: Models are trained from scratch on extensive datasets focused on Indian languages, including a corpus of over 2 trillion authentic Indian language tokens.
Development Tools: Utilizes NVIDIA NeMo framework and NVIDIA Megatron-LM for pretraining, and NeMo-RL for post-training workflows, demonstrating a full-stack hardware-software co-design approach with NVIDIA.
Open-Source Availability: Sarvam-30B and Sarvam-105B were open-sourced under the Apache License 2.0, with model weights made available on Hugging Face and AIKosh.
Other Key Models/Products:
- Sarvam-1 (2B parameters): An earlier model optimized for Indian languages, trained on 2 trillion tokens from 10 Indic languages and synthetic data.
- Sarvam Vision: An advanced OCR and document intelligence model built on a State-Space Model (SSM) backbone, capable of reading high-resolution pages without downsampling and processing handwritten and Indian-language records.
- Bulbul V3: A highly expressive text-to-speech AI system supporting over 30 voices in 11 Indian languages.
- Saaras: A dedicated speech-to-text model.
- Sarvam Translate / Mayura: Models designed for translation across various Indian languages.
Deployment & Infrastructure: Offers REST APIs, SDKs, cloud-based inference, VPC deployment, and on-premise installations. Emphasizes custom fine-tuning, controlled data residency, and infrastructure-level integration. Sovereign compute ensures AI inference occurs on Indian soil, hosted with partners like Yotta in Navi Mumbai.

🔮 Future ImplicationsAI analysis grounded in cited sources

Sarvam AI's strategic partnership with HCLTech will significantly accelerate its penetration into enterprise and government sectors, both in India and potentially globally.

HCLTech's substantial investment and global client relationships provide a direct channel for Sarvam's sovereign AI solutions to reach a broader market, particularly in critical sectors like banking, insurance, and defense.

The company will likely prioritize the development of advanced agentic AI, coding, and cybersecurity models, leveraging the new funding to expand its compute capacity and research efforts.

The Series B funding is specifically earmarked for continued research on its next frontier model for agentic, coding, and cybersecurity use-cases, along with access to compute at scale.

Sarvam AI's focus on open-source models and a full-stack approach will foster a robust 'Made in India' AI ecosystem, reducing reliance on foreign AI technologies for critical national applications.

By open-sourcing models and building a full-stack platform with a 'sovereign by design' philosophy, Sarvam AI aims to empower Indian developers and enterprises to build AI solutions that are culturally resonant and maintain data sovereignty.

⏳ Timeline

2023-08

Sarvam AI founded by Vivek Raghavan and Pratyush Kumar.

2023-12

Raised approximately $41 million in a combined seed and Series A funding round.

2025-03

Collaborated with UIDAI to integrate AI-based voice interactions into Aadhaar services.

2025-04

Selected by MeitY to develop an indigenous foundational model under the IndiaAI Mission.

2026-02

Released open-source foundational models Sarvam-30B and Sarvam-105B at the India AI Impact Summit.

2026-06-15

Achieved unicorn status with $234 million in Series B funding led by HCLTech, reaching a $1.5 billion valuation.

Sarvam AI becomes India's newest unicorn with $234m funding

⚡ 30-Second TL;DR

🧠 Deep Insight

🔑 Enhanced Key Takeaways

🛠️ Technical Deep Dive

🔮 Future ImplicationsAI analysis grounded in cited sources

⏳ Timeline

📎 Sources (25)

👉Related Updates

Gravity SMTP flaw exposes API keys on 100,000 sites

Harvard Business Review warns AI ‘workslop’ is rotting companies

Microsoft finds USB worm stealing cryptocurrency via Tor

AirPods Pro 3 heart rate sensor accuracy tested