Is there a better small model than Qwen3.5 4B for a fast local AI assistant?

A user is seeking a small, fast local AI assistant model that significantly upgrades Qwen3.5 4B, which they currently use and achieve 40-50 TPS with. They are looking for a model with a similar memory/compute footprint that offers substantial improvements in general conversation, multilingual ability, and tool calling, as they are not interested in marginal differences.