Bengaluru-based Sarvam AI closed a $200 million Series B on Friday, led by SoftBank Vision Fund 3, and simultaneously released BharatGPT-3 — a 70-billion-parameter language model covering 22 Indian languages, including seven that had no representation in any commercial LLM prior to this release. The round brings Sarvam’s total funding to $247 million and values the company at approximately $1.4 billion.
What BharatGPT-3 Actually Does
The model is trained on 4.2 trillion tokens of multilingual Indian-language data — a corpus Sarvam says took 18 months to assemble from digitized government records, regional news archives, academic publications, and curated web crawls across Indic scripts. Supported languages include Devanagari-script variants (Hindi, Marathi, Nepali, Sanskrit), Dravidian languages (Tamil, Telugu, Kannada, Malayalam), and Eastern Indic scripts (Bengali, Odia, Assamese) alongside Punjabi, Gujarati, and 10 additional regional languages.
On the IndicBench-2026 evaluation suite — a standardized benchmark across reading comprehension, translation, summarization, and reasoning tasks — BharatGPT-3 outperforms GPT-4o on 18 of 22 tested languages. GPT-4o retains an edge in Hindi and Bengali, where training data volume is largest across Western AI providers. The seven newly covered languages — including Maithili, Dogri, Bodo, and Santhali — have no prior commercial LLM benchmarks to compare against.
Inference via Sarvam’s API is priced at ₹0.85 per 1,000 tokens, approximately $0.01 at current exchange rates, targeting cost parity with Hindi-only alternatives and undercutting OpenAI’s comparable tier by roughly 85%.
Government Alignment and Real Deployments
India’s Ministry of Electronics and IT has formally designated Sarvam a “Strategic AI Partner” under the IndiaAI Mission, the government’s ₹10,372 crore ($1.25B) AI infrastructure initiative announced in the Union Budget 2025–26. The designation comes with preferential procurement status for government digital services deployments.
That isn’t purely aspirational: Sarvam currently powers chatbot infrastructure across 14 million monthly active users in government-facing applications, including tax filing guidance for India’s Income Tax Portal, health scheme eligibility queries under Ayushman Bharat, and judicial document translation for seven high courts. The judicial translation use case alone covers an estimated backlog of 4.8 million untranslated case documents, according to the Supreme Court e-Committee’s 2025 annual report.
SoftBank’s participation is notable given the Vision Fund’s historically mixed record in Indian tech. Lightspeed India and Peak XV (formerly Sequoia India) — both existing investors — participated in the round, signaling continued confidence from the domestic VC ecosystem.
Why Linguistic AI Coverage Matters at Scale
Only 7% of the world’s approximately 7,100 languages have any representation in current commercial LLMs, according to data compiled in the AI Index 2026 report (Stanford HAI). The mismatch is most acute in South and Southeast Asia, where hundreds of millions of speakers interact with government services, financial institutions, and healthcare systems using languages that existing AI tools treat as out-of-scope.
OpenAI, Google, and Meta collectively cover 95+ languages, but coverage depth skews heavily toward languages spoken in high-GDP markets. BharatGPT-3’s architecture — fine-tuned on structured government and judicial text rather than general web crawl data — also reflects a different optimization target: accuracy in formal administrative language rather than conversational breadth.
Sarvam’s raise follows Gulf states’ $50 billion commitments to sovereign AI infrastructure announced earlier this week, reinforcing a global pattern of large economies building domestically calibrated AI capabilities rather than depending exclusively on US hyperscaler API offerings. The next test for BharatGPT-3 is whether enterprise and financial services customers — India’s banking sector serves 500 million active account holders — adopt it at sufficient scale to justify the $200 million bet.
Discussion
Sign in to join the discussion.
No comments yet. Be the first to share your thoughts.