Xiaomi unveils HySparse2 architecture to boost AI model efficiency
Chinese tech company Xiaomi has unveiled its new HySparse2 core architecture for its MiMo-V3 model. The new system significantly enhances AI models' ability to process and retrieve information from long contexts and complex tasks.
Chinese technology firm IT Home (Xiaomi) has introduced its new HySparse2 core architecture for its MiMo-V3 model. The architecture is designed to enhance the performance and efficiency of AI agents tackling long-context, multi-turn tasks, particularly when processing extensive data like web pages, documents, and code.
HySparse2 aims to reduce computational costs and memory usage while improving the accuracy of information retrieval from growing task histories. The company states the architecture allows models to more efficiently process large datasets, such as web pages and files, and extract critical information from complex conversations and task sequences.
The new architecture builds upon the previous HySparse model by upgrading KV Cache sharing and sparse selection mechanisms. It features a two-level KV sharing approach and token-level sparse selection, enabling more precise information retrieval without a substantial increase in computational power. The model can also continuously monitor the latest information, crucial for multi-step processes.
Xiaomi reports that HySparse2 significantly reduces Prefill computations and KV Cache requirements compared to earlier models. In tests with one million tokens, Prefill computation was reduced to one-fifth, and KV Cache usage dropped to 2.7 GB. This substantially lowers computational load and enables faster responses, which is particularly important for AI applications requiring long contexts.