I got Qwen3.5 35B A3B (~21 GB / 35B MoE) running on an RTX 2050 with just 4 GB VRAM and 16gb ram. Can token generation be improved further?

TL;DR: We modified llama.cpp so that Qwen3.5-35B-A3B (Q4_K_M, ~21 GB GGUF) runs on a laptop RTX 2050 (4 GB VRAM) at around 1.2 tok/s , exposing an OpenAI-compatible API that's actually usable for agentic coding . so I'm here looking for ideas to improve generation speed. Also is there any opensourced MOE model that loads experts for per input prompt and not per token? Non-Techie Description : Think of the AI model as a huge library containing hundreds of books (called experts ). A normal setup tries to plac