Hybrid Models

Researchers introduce HeadWiseKV to compress hybrid language model caches: Researchers have introduced HeadWiseKV, a training-free framework designed to compress the residual global key-value (KV) caches of hybrid language models. It assigns each physical KV head a static, multilevel history window to make cache demand predictable before serving.

LLM • arXiv

Perplexity launches hybrid compute to process sensitive data locally: Perplexity has launched hybrid compute for its Computer agentic platform, allowing a single AI agent to split tasks between cloud-based frontier models and local open-weight models on Apple silicon Macs. This system routes confidential data to the local machine without restarting the job or losing context.

Perplexity • VentureBeat