Challenges in Adoption for Indian LLMs

Can India build its own foundational LLMs? Yes
But who is using them? How much is their adoption?

To find answers to these questions, I’ve sourced publicly available data from various sources as below:

1️⃣ Number of Downloads on Hugging Face
Hugging Face is the de-facto platform for developers to download AI models and datasets. I’ve considered the number of downloads (as a proxy for usage and adoption) of leading, open-source LLMs from USA (from Meta), China (from DeepSeek AI & Alibaba Cloud), and India (from Sarvam & Krutrim, as the two most well capitalized Generative AI startups). The data shows that in the same time period of the last one month from today:

- US: LLama’s 3.2-1B & 3.1-8B-instruct were downloaded ~11M & ~6M times
- China:  DeepSeek-R1 & Qwen2-VL-7B-instruct were downloaded ~4M & 1.5M times
- India: Sarvam-1 & Krutrim-2-instruct (built on top of Mistral-NeMo 12B) were downloaded ~5k and ~1k times

👉 These numbers show that the adoption of our leading LLMs is 3 to 4 orders of magnitude less than the most popular LLMs from China and USA respectively. The absolute numbers might be slightly different as these LLMs are also available as APIs, on cloud platforms etc. but the overall trend may not be that different.

2️⃣ Number of forks of Github repositories
Forking of Github repos represents a stronger sign of adoption by the developer community, and here also the picture is similar:

- meta-llama has been forked ~9700 times
- DeepSeek-v3 has been forked ~13800 times
- DeepSeek-R1 has been forked ~10000 times
- Qwen-VL has been forked 400 times
- Krutrim-2-12B has been forked 6 times
- Sarvam doesn’t have a dedicated repo for Sarvam-1

3️⃣ Listing in LLM Marketplaces
Customer-centric LLM marketplaces like AWS BedRock also provide an indication of customer usage & adoption. While Meta’s LLama and DeepSeek-R1 models are supported, none of India’s LLMs are available.

4️⃣ Support from LLM inference engines
LLM Inference engines like vLLM also provide signals about LLM adoption for production use cases. vllm currently supports Llama and Qwen models but again no Indian LLMs yet.

5️⃣ Conclusions
Overall, the analysis indicates that Indian LLMs do not currently receive significant user interest and therefore their impact is far less than top, global LLMs.

Our LLMs likely have a competitive advantage for domestic use cases focused on speech and language e.g. translation, document analysis, speech recognition etc. The market size of our domestic use cases may not be big enough to justify investment by global companies, but it clearly represents an area where indigenous LLM builders can distinguish themselves.

Following my previous post on the poor trajectory of India’s AI research record at top AI conferences, these data further show that we are far from the cutting-edge of AI research and a lot of work needs to be done to raise the bar in terms of global adoption and impact.