IIM Lucknow Alumnus Builds Open AI Models in 13 Days Through Udgam Labs
Jitesh Prasad, a physics graduate from the University of North Bengal and an alumnus of IIM Lucknow, is drawing attention to a different approach to artificial intelligence development. Through his small venture, Udgam Labs, Prasad has released a series of AI models and research projects focused on building useful systems with limited computing resources.
In October 2026, Udgam Labs reported a rapid sequence of releases covering areas such as speech recognition, language modelling, reasoning, mathematics and image-related AI. The work also included a public benchmark intended to help compare smaller AI models.
Rather than focusing only on model size, the projects examine how far smaller systems can go when training costs, computing capacity and development time are constrained.
From Physics and Management to AI Research
Prasad's academic background combines physics and management. He studied physics at the University of North Bengal before completing an MBA at IIM Lucknow.
Physics involves testing assumptions against evidence, while management education often focuses on costs, execution and scalability. These perspectives provide a useful way to understand the approach described in Udgam Labs' projects.
The lab has emphasised publishing information about its experiments, including training methods, evaluations and limitations. This approach allows other researchers and developers to examine the results rather than relying solely on performance claims.
One of the lab's English-language model experiments, for example, did not reach its stated performance target. According to the supplied account, Udgam Labs still published the result alongside the training material and evaluation details.
This emphasis on documenting unsuccessful experiments is relevant to open AI research, where transparent reporting can help others understand which methods work and where further improvements are needed.
Laghu: A Small Language Model Built on a Limited Budget
One of Udgam Labs' projects is Laghu, a chat-oriented language model with 152 million parameters.
According to the project's reported figures, its main training run cost $146, bringing the total project cost to approximately $194. Training took 13 hours using eight NVIDIA A100 GPUs.
The project made its code, model checkpoints, training figures and evaluation results publicly available, allowing other developers to inspect the work and assess its performance.
Laghu illustrates the lab's central research question: how much capability can be developed in a smaller AI model without relying on the enormous budgets and computing infrastructure associated with leading large-scale systems?
Its reported training cost provides one measure of the experiment, although model quality must also be assessed through independent testing, task performance and practical use.
Shruti: Compact Speech Recognition for Multiple Languages
Another project, Shruti, focuses on speech-to-text technology. The model has a reported size of 29.5 MB and was tested on Hindi, Hinglish and English speech.
The reported comparisons show different results across languages. On the Hindi evaluation material, Shruti recorded fewer errors than Whisper large-v3. On English speech, however, Whisper large-v3 performed better.
These results suggest that the performance of a speech recognition system can vary according to language and evaluation data. A compact model may be useful in specific settings, but comparisons need to account for the test dataset, accuracy measurements and intended application.
Shruti's small size is also relevant for developers exploring speech recognition in environments where storage, computing power or deployment resources are limited.
Gargi: Testing the Limits of Smaller Reasoning Models
Udgam Labs also developed Gargi, a language model with 384 million parameters trained on approximately 17 billion tokens.
The reported comparison placed Gargi against Qwen2.5-0.5B, a model with 500 million parameters that was trained on roughly 18 trillion tokens.
According to the supplied account, Gargi came within two percentage points of Qwen2.5-0.5B on several reasoning benchmarks. However, it performed considerably worse on knowledge-based evaluations.
The results highlight an important distinction in AI development: a model's performance can vary considerably depending on the task being tested. Results on selected reasoning benchmarks do not necessarily translate into comparable performance across general knowledge, language understanding or other applications.
Gargi's evaluation therefore provides a case study in the trade-offs involved in developing smaller models with more limited training resources.
Why Open AI Models Matter
Open model releases can give researchers, students and independent developers an opportunity to examine how AI systems are trained and evaluated. Publicly available code, checkpoints and documentation can also help others reproduce experiments, identify weaknesses and test alternative approaches.
For smaller research teams, this can be particularly valuable. They may not have access to the computing budgets available to major technology companies, but carefully scoped experiments can still contribute to discussions about model efficiency and performance.
However, low training costs and compact model sizes do not automatically establish that a model is better than a larger alternative. Performance depends on the task, quality of training data, evaluation methodology and deployment requirements.
Independent reproduction of benchmark results and clear documentation of model limitations remain important when assessing the practical value of these projects.
A Different Approach to AI Development
Udgam Labs' reported releases place the emphasis on experimentation, cost transparency and the evaluation of smaller AI systems. Prasad's work brings together a background in physics, management education and hands-on model development.
The projects also show why it is important to examine more than a headline performance score. Laghu demonstrates an effort to document the cost of language-model training, Shruti highlights differences in speech recognition across languages, and Gargi illustrates how results can vary between reasoning and knowledge benchmarks.
Together, these experiments offer a perspective on AI research outside the largest technology laboratories. Their broader significance will depend on reproducibility, further evaluation and how effectively the models perform in real-world applications.
Comments