We shrank Whisper large-v3 by 68%, below the size of turbo – Tileward (tileward.com)

Tileward has announced a significant advancement in AI model efficiency by shrinking the Whisper large-v3 model size by 68%, now lighter than existing Turbo models. This compression allows businesses to serve compressed model weights on their hardware, drastically cutting down on inference costs by only sending relevant parts of a conversation needed to answer the next question. This innovation not only optimizes resource use by preventing unnecessary data from being processed, but also enhances compliance by allowing organizations to enforce their data policies upfront. The implications for the AI and machine learning community are profound, as Tileward effectively addresses long-standing challenges in managing context and resource allocation in multi-agent systems. This approach minimizes memory and context capacity limitations while still ensuring that policy governance is maintained. The prototype demonstrated up to 19 times less context used while achieving a 55% answer rate, which is comparable to the full operational context performance. Tileward’s model operates on a flexible pay-as-you-go model, making it accessible for providers looking to efficiently integrate AI into their infrastructure while staying within budget and regulatory frameworks.