Indian startup NeuroDX's 240M-parameter model Sphota slightly outperforms Liquid AI on about 11 benchmarks
Indian startup NeuroDX claims that its 240M-parameter model Sphota scores an average of 52.4% on about 11 benchmarks, slightly above Liquid AI's 52.1%, and claims it uses 50 times less training data and is 4.5 times more memory-efficient. These results have not yet been verified by independent public evaluations, and the company is developing versions with 1 billion to 2 billion parameters, targeting release around December. NeuroDX will also be renamed Manaas AI, and the Sphota architecture will remain closed-source for the time being.
#Model releases#English text is being prepared.#Open-source ecosystem