Increasing the --parallelism flag can potentially increase the throughput of tabby, but it will not have an impact on reducing latency for single-user setups. To reduce latency, you can try using a smaller model size or using a better device such as a GPU.