Krutrim Now Hosts DeepSeek Models On Its Cloud

Bhavish Aggarwal’s AI unicorn Krutrim is now hosting open source AI models of Chinese GenAI company DeepSeek on its cloud platform. 

“Krutrim has accelerated efforts to develop world class AI. As a first step, our cloud now has DeepSeek models live, hosted on Indian servers. Pricing lowest in the world,” Aggarwal said in a post on social media platform X.  

Further hinting on the open source efforts, Aggarwal said that the details on its “AI lab, state of the art (SOTA) model and research progress, open source drops on 4th Feb!”

As of now, five models of DeepSeek are available on the platform, ranging from 8 Bn tokens to 70 Bn trained models. 

Apart from the main model DeepSeek R1, which came into the news for beating OpenAI’s GPT-4o in many benchmarks, Krutrim also hosts distilled versions like DeepSeek-R1-Distill-Llama-8B, DeepSeek-R1-Distill-Qwen-14B, DeepSeek-R1-Distill-Qwen-32B,  DeepSeek-R1-Distill-Llama-70B. 

Overall, DeepSeek has launched six distilled models based on Alibaba and Meta’s open source models Qwen and Llama, respectively. 

Model distillation is a process of transferring knowledge from a large model to a smaller model. DeepSeek says that R1-Distill models are fine-tuned based on open-source models, using samples generated by DeepSeek-R1. 

Notably, the R1-Distill-Llama-8B model on Krutrim’s cloud only costs about INR 10 per 1 Mn tokens generated. The costliest DeepSeek model costs about INR 60 per 1 Mn tokens. 

For context, tokens are a fundamental unit of text that LLMs generate. 

Founded in April 2023, Krutrim offers a foundational generative AI model, GPU-as-a-service, model-as-a-service, along with different no code platforms. It became a unicorn in January last year after raising $50 Mn in a round led by a clutch of investors including Matrix Partners India.

The update comes as the recent launch of DeepSeek’s R1 model has shaken the tech industry worldwide, with its model costing about 95% cheaper than OpenAI’s GPT-4 model. Further, while GPT-4 training cost was about $100 Mn, DeepSeek claims that its model cost only under $6 Mn. 

After DeepSeek’s launch, Nvidia’s stock tumbled– along with most of the US tech stocks. 

Meanwhile, IT minister Ashwini Vaishnaw recently said that India is also planning to build a domestic LLM as part of the INR 10,037 Cr IndiaAI Mission. Along with this, he also said 

that India will also host DeepSeek on local servers to address the privacy concerns regarding the cross-border data transfer. 

On the same day, IndiaAI Mission sought proposals for Indian foundational models, which will be shortlisted on factors like innovativeness of the approach, scalability and sustainability, financial viability, ethical considerations, among others.

Under the plan, the Mission will offer “milestone-based disbursements” in the form of a direct grant and compute credits for AI Compute.