Farang has built a new LLM architecture that internally comprehends a complete response before translating it into text, an approach the company says mimics human thinking and differs from token-by-token prediction. The architecture performs internal reasoning through non-textual mechanisms, which Farang claims yields more coherent answers while sharply reducing computational demands—about twenty-five times less compute in current testing. The company is focusing initially on specialised domain models where existing assistants underperform, with a first product focus on React programming to deliver more optimized and iterative code. Farang also targets niche medical fields and internal company tools, and its models can be deployed on-premises to preserve data privacy for healthcare, legal, and financial firms. Founder Emil Romanus says the plan is to prove the architecture in specialised domains and then scale toward broader AI applications. The recent funding will be used to scale proof-of-concept models and buy compute to train and fine-tune models for programming and medical use cases.

Total raised

$2M

Funding rounds

1

Latest round

Seed

Latest activity

Aug 2025