Bengaluru-based startup Sarvam AI has been selected by the Government of India, under the IndiaAI Mission, to build a sovereign foundational large language model (LLM)* fully developed in India. The project is expected to be completed within six months.
Key Details & Commitments
- The model will have about 70 billion parameters, a size aimed at enabling strong reasoning, multilingual fluency, and voice-based interaction.
- It will be built from scratch using local infrastructure, delivered by Indian talent, and optimized for Indian languages and cultural context.
- As part of the deal, Sarvam AI will receive ~4,000 Nvidia H100 GPUs over six months to support the training and development
- There will be three model variants planned:
- Large – for advanced reasoning and generation tasks.
- Small – optimized for real-time, interactive applications.
- Edge – compact models for on-device deployment in phones/IoT etc.
Why It’s Important
- Strategic autonomy: India’s effort to build its own foundation model means less dependence on foreign AI providers.
- Language & culture adaptation: The model is intended to be fluent in Indian languages and voice input, helping accessibility and relevance for local users.
- Building compute infrastructure capacity: Allocating thousands of GPUs, supporting domestic data centers, will also strengthen India’s AI infrastructure.
- Competing globally: Officials believe that this LLM can be competitive with international models in performance.
Challenges Ahead
- Data diversity and quality: Gathering clean, well-labeled data across many Indian languages and dialects will be demanding.
- Bias, fairness & ethical concerns: Ensuring the model doesn’t amplify biases (caste, religion, gender, etc.) is crucial.
- Compute & cost constraints: Even with 4,000 GPUs, training a 70B parameter model is expensive, complex, and energy intensive. Infrastructure stability, efficiency, and optimization will be key.
- Benchmarking vs global leaders: The expectations are high; global LLMs have had years of research, massive datasets, and large teams. Matching up probabilistically in performance may be tough.
Timeline & What to Watch For
| Milestone | What to Expect |
|---|---|
| Project kick-off | Training reportedly begun around June with ~1,500 GPUs, with more to be added. |
| Full deployment | Six-month target from announcement (April 2025 → around Oct/Nov 2025) for readiness of the foundational model. |
| Open sourcing / accessibility | Earlier reports conflicted: some say the model will be proprietary; others say open source. Need clarity. |
| Use cases | Expected applications: voice-based assistants, multilingual platforms, enterprise tools, apps adapted for Indian languages at scale. mint |
