The Visual Computing Lab at IISc Bangalore was working on a problem that remains difficult for most commercial vision models: understanding hand-drawn sketches. The research led to O3SLM, an open-weight multimodal model released in 7B and 13B versions, accepted at AAAI 2026, and benchmarked at state-of-the-art performance across multiple sketch datasets.
The biggest obstacle was not simply training a larger model; it was sustaining the full research workflow.
The IISc team had to:
● generate 32+ million sketch-image pairs from scratch because suitable training data did not exist,
● train a memory-intensive multimodal architecture that processes sketches, photographs, and language together,
● and run multiple training and evaluation cycles to reach state-of-the-art performance.
These requirements created a need for dedicated high-memory GPU capacity, sustained throughput, and the freedom to iterate without compromising model size or experimental scope.
Neysa supported the project through Velocis Bare Metal, its dedicated AI infrastructure platform, providing:
● Dedicated high-memory GPU nodes
● Full-scale multimodal training capacity
● Sustained throughput for large-scale data generation
● Rapid experimentation and repeat training cycles
● Predictable scaling for research workloads
The entire O3SLM development pipeline - from dataset generation to fine-tuning and benchmark evaluation - was run on Neysa Velocis Bare Metal.
Key outcomes include:
● 32M+ sketches generated for training
● 7B and 13B open-weight model releases
● State-of-the-art performance across four sketch benchmarks
● AAAI 2026 acceptance
It would be great if you would consider exploring a story on how dedicated AI infrastructure is supporting advanced research workloads that would otherwise face constraints around memory, iteration speed, and scaling.
I am attaching the full case study for reference. We can arrange conversations with Neysa’s AI infrastructure team as well as a spokesperson from IISc Bangalore.
See What’s Next in Tech With the Fast Forward Newsletter
Tweets From @varindiamag
Nothing to see here - yet
When they Tweet, their Tweets will show up here.




