Editorial illustration for Manas AI Rebrands After Launching 240M-Parameter Model
Indian Startup's 240M Model Beats Liquid AI
Manas AI Rebrands After Launching 240M-Parameter Model
A 240-million-parameter model built by an Indian startup is now outscoring a comparable system from Liquid AI, and it got there with a fraction of the usual training data. NeuroDX, backed by the IndiaAI Mission, says its model Sphota beat Liquid AI's entry 52.4% to 52.1% across roughly 11 benchmarks. CEO Siddharth Panwar claims the gap in resources is the real story: Sphota reportedly needed 50 times less training data and runs with 4.5 times better memory efficiency. Those numbers come from NeuroDX's own reported results, not an independent public evaluation, so they're preliminary.
What makes the claim interesting isn't just the scorecard. NeuroDX didn't start out chasing efficient small models at all. The company's roots are in brain research, building AI systems to interpret EEG signals, the messy electrical readouts of scalp-recorded brain activity that are far harder to parse than an MRI.
That work, done with backing from the Ministry, pushed NeuroDX toward an architecture that tries to capture meaning directly instead of leaning on next-word prediction, the approach most large language models are built on. Sphota is the first language model to come out of that detour.
Notably, NeuroDX CEO Siddharth Panwar claims Sphota achieves comparable performance with 50 times less training data and offers 4.5 times better memory efficiency.
Why this matters A 0.3-point edge on 11 benchmarks is not a knockout blow, and founders chasing foundation-model status should notice how thin that margin is before treating Sphota as proof of a new training paradigm. What's genuinely worth watching is the pivot itself: a company born from brain research and backed by the IndiaAI Mission is now rebranding as Manas AI to chase general-purpose intelligence, which says as much about where government-backed AI money wants to go in India as it does about the model's actual capabilities. For researchers, the claim of matching Liquid AI's performance with less training data deserves scrutiny, not applause, until the benchmark methodology and data sourcing get independently checked.
For developers, a 240-million-parameter model that's even competitive at this scale is worth testing on your own workloads, cheap compute matters more than marginal leaderboard wins. The real story here isn't the score, it's a healthcare startup betting its whole identity on becoming a frontier lab. Whether that bet pays off depends on what Manas AI ships next, not what it just rebranded into.
Related Reading
Research & Benchmarks Microsoft AI Chief: "If It Isn't Safe We Shouldn't Build It
Research & Benchmarks Google's FACTS benchmark shows 70% factuality ceiling across four tests
Research & Benchmarks Nvidia's DLSS 4.5 beta adds 6x Multi Frame Generation for RTX 50 GPUs
From Archives Adobe Tests AI Agent EditAgent Using Editing Tools Like Photoshop Pros
From Archives Study: Chinese AI models shift pro-government when prompted in Chinese
Common Questions Answered
How does Sphota's performance compare to Liquid AI's model on benchmarks?
Sphota, NeuroDX's 240-million-parameter model, outscored Liquid AI's comparable system 52.4% to 52.1% across roughly 11 benchmarks. While the margin is relatively thin at 0.3 points, it demonstrates competitive performance in a direct comparison between the two systems.
What are the key efficiency advantages of Sphota according to CEO Siddharth Panwar?
Sphota achieves its competitive performance with 50 times less training data and offers 4.5 times better memory efficiency compared to Liquid AI's model. These resource advantages represent a significant gap in the efficiency of the training and deployment processes between the two systems.
Why is Manas AI rebranding and what does it signal about India's AI strategy?
NeuroDX, originally founded on brain research and backed by the IndiaAI Mission, is rebranding as Manas AI to pursue general-purpose intelligence development. This pivot reflects the direction that government-backed AI funding in India is moving toward, prioritizing frontier AI capabilities over specialized applications.
What backing does the Sphota model have and how does it relate to India's AI landscape?
Sphota is developed by NeuroDX, which is backed by the IndiaAI Mission, positioning it within India's government-supported AI research ecosystem. This backing demonstrates India's commitment to developing competitive frontier AI models domestically.
Further Reading
- Papers with Code Benchmarks - Papers with Code
- Chatbot Arena Leaderboard - LMSYS
