BloombergGPT (50 billion parameters, a 363-billion-token proprietary corpus)
was beaten by GPT-4 with zero financial pretraining at all. Seven medical
LLMs tested against their own base models won 12.1% of the time and were
significantly worse 38.2% of the time. Krutrim spent ₹10,000 crore and left
model-building inside thirty months.
And the counter-evidence is what makes the corrected thesis work: the same
Bloomberg paper shows small fine-tuned models beating GPT-4 by 26 points on
domain-specific entity recognition, and a 2026 industrial study found
retrieval-augmented generation the most cost-efficient adaptation method,
with 3-billion-parameter open models matching frontier ones.
Small models win where the data is proprietary and the task is
narrow. Frontier models win at language. We build the first and rent the
second: training small vision and tabular models, renting the language
layer from whoever is best that quarter, and owning the data access.