When Abhinav Anand first discovered ChatGPT nearly two and a half years ago, building an artificial intelligence model was not on his mind. The 19-year-old from Bihar was simply curious about the technology and wanted to experiment with it.
Today, the Class 12 student is building a 5.82-billion-parameter multimodal AI model called “ArcleIntelligence” largely on his own, without investors, a team, or a formal computer science degree.
His story caught eyeballs after he posted on Reddit detailing how he spent nearly $11,560 (around ₹9.6 lakh) on GPU compute, cloud credits, and training infrastructure to build the model.
“Posting here because this community understands what it means to build something real in India with nothing,” he wrote.
The post quickly went viral among developers and AI enthusiasts, particularly because the project emerged not from a startup ecosystem or research lab but from a middle-class household in Bihar.
Failing Fast
Anand says his journey into AI began with repeated failed experiments.
Before working on ArcleIntelligence, he attempted to build a YouTube analytics platform, a voice assistant, and an offline AI assistant. None of the projects succeeded.
“Every failed project taught me something practical. I slowly started understanding how models work, how training works, and how AI systems are actually built,” he tells AIM.
One of his early breakthroughs came when he trained a text-to-video AI model entirely on his personal laptop. The project later caught cloud-based platform Lightning AI's attention, which invited him to publish it as an official studio template on its platform.
“That was the first moment where I felt maybe I could build something bigger,” he says.
Building a Multimodal AI Model
Soon, Anand began building ArcleIntelligence, a multimodal AI system capable of handling text, images, documents, audio, and video inputs. The model, according to him, can generate text responses, images, and speech outputs while supporting a context window of more than two million tokens.
He felt the need to develop an AI model after he found a gap in edge AI computing.
“When I was looking for models for my AI assistant on the internet, I came to know that there is no such model for my specific use case, like a type of Gemini Omni model. There are many Omni-like models, but none that can run on your own device,” the teenager shares.
Anand has used mixed-precision training with 8-bit weights to train the model. He has largely trained the model himself using open-source datasets, including sourcing 8 trillion tokens from Hugging Face at 5.6 million tokens per second. But he has also pre-trained it on an image generation model as well as on OpenAI’s open-source automatic speech recognition model Whisper.
He is building the model on SSM architecture. The model’s context window is 2 million tokens.
Anand claims that in private testing, the model scored 93.45 on OmniDocBench V1.5, a benchmark for evaluating document-understanding capabilities in AI systems. The score, which could not be independently verified, drew attention online due to comparisons with systems developed by companies such as OpenAI, Google and Alibaba.
The teenager says much of the project relied on personal savings of ₹1.2 lakh, free student developer resources, and over 6,000 compute credits he won after participating in OpenAI Parameter Golf machine learning challenge.
His family currently have little knowledge of what he is building, but he plans to tell them eventually when the model is complete.
“My father is a government officer, and my mother is a housewife,” he says. “For a middle-class family in Bihar, spending nearly ₹10 lakh on a computer is definitely not a small thing.”
To continue training the model, he used cloud infrastructure credits from cloud computing platforms RunPod and DigitalOcean, and the GitHub Student Developer Pack.
Why Open Source
For Anand, the project is not just about building another AI model. He believes India needs more independent AI development that is accessible to students and researchers. “The West has OpenAI. The east has DeepSeek. India also deserves something built by Indians and accessible to everyone.”
He plans to eventually release the model publicly on platforms like GitHub and Hugging Face under an open-source licence. The goal is to allow developers to freely use, fine-tune, and build upon the model without restrictions. The model is currently in the pre-training stage.
He is currently attempting to raise nearly $35,000 to complete the remaining stages of development.
Your reaction
Discussion
What's your take on this story?
No comments yet — be the first to share your view.