User runs Qwen3.6 35B A3B on 20GB VRAM + 64GB RAM for local coding

A Reddit user shares their experience running the Qwen3.6 35B A3B model locally on a laptop equipped with a 20GB 7900XT eGPU and 64GB DDR5 RAM.

• The model is used for local pi-coding-agent tasks.
• It utilizes CPU offload (MoE) to fit high quantizations, such as Q8 at 100k context or ud-q6kl at full 262k context.
• Inference speed is described as slower but usable for outputs exceeding 20 tokens.

The user seeks recommendations for other Mixture of Experts (MoE) models suitable for this specific hardware configuration.