VibeStudio claims major efficiency breakthrough with pruned open-source LLM

Deeptech startup VibeStudio has announced a significant advance in efforts to make artificial intelligence (AI) models lighter and more energy-efficient. 

VibeStudio, incubated at Immersive Technology and Entrepreneurship Labs in Chennai, said it has achieved a 55% reduction in the size of the MiniMax M2 open-source large language model (LLM) while retaining close to 80% of its reasoning and coding capability.

It has released the results as open source on HuggingFace and is intended to support the large-scale deployment of AI coding tools on modest hardware.

The team, led by founder Arjun Reddy, developed a structured pruning process called THRIFT, short for Targeted Hierarchical Reduction for Inference and Fine Tuning.

According to the company, the method analyses the model layer by layer and removes redundant pathways or inactive parameters. Each pruning stage is followed by fine-tuning to stabilise performance. This approach avoids the severe degradation that can occur when models are compressed in a single sweeping step, it noted.

The push for smaller models reflects a growing shift in parts of the Indian AI ecosystem. While global labs continue to expand model sizes, Indian firms have increasingly concentrated on efficiency and accessibility. 

Startups and initiatives like Sarvam AI, AI4Bharat, CoRover AI, and Gnani AI have explored compact models for multilingual tasks, speech systems, and economically deployable AI assistants. 

Much of this work has targeted resource-constrained environments, including Indian universities, public sector systems, and mid-sized enterprises that lack access to large GPU clusters. 

VibeStudio’s effort fits squarely into this pattern, aiming to prove that smaller systems can deliver practical performance without prohibitive costs.

The company said its pruned M2 model performs well for software engineering tasks inside VibeStudio’s agentic development environment. It supports structured code suggestions, multi-file reasoning, refactoring, and automated workflows. 

These capabilities are key to the company’s ambition to offer a secure on-premise alternative to cloud-dependent coding assistants. The startup reports more than 150,000 downloads across its open-source releases, which it attributes to what it calls "a focus on disciplined engineering rather than marketing hype".

VibeStudio is also developing proprietary models for enterprise customers, including an 8-billion-parameter dense model and a 32-billion-parameter mixture-of-experts system.

These are designed for organisations that require predictable behaviour, data privacy, and in-house deployment. Although these larger models remain closed source, the company argued that open releases and private deployments can coexist in a hybrid strategy that addresses different market needs.

Indian companies have increasingly turned to such hybrid approaches. Enterprises in the finance, logistics, and telecommunications sectors have experimented with compact in-house models while relying on open releases for experimentation and training.

This trend has grown as GPU shortages and rising electricity tariffs have encouraged firms to seek cost-efficient alternatives.

ITEL, the not-for-profit organisation that incubates VibeStudio, said the work aligns with its broader mission to strengthen India’s capacity to build sovereign AI systems. The organisation has also launched the Vikram Sarabhai AI Labs to accelerate domestic research and support startups working in high-impact areas.