Qwen3 Fits 64GB RAM, 8GB VRAM

The answer is less about maximum parameter count than where the bottleneck lands. On this class of machine, 4-bit quants and a lean loader matter more than chasing the biggest checkpoint you can technically boot.

• –Qwen3.6-35B-A3B gets the clearest support; its sparse MoE design makes it a better fit than a dense 35B model for mixed RAM/VRAM setups.
• –CPU offload extends capacity, but it taxes throughput hard; commenters specifically warn that IQ-style quants get much less attractive once experts spill onto the CPU.
• –For interactive chat and coding, a smaller dense model can feel better than a larger offloaded one because response time usually matters more than raw size.
• –Long-context work is where 64GB RAM helps most, but once you lean on system memory too heavily, the user experience shifts from “local powerhouse” to “patient batch job.”
• –GGUF/llama.cpp-style stacks, including Ollama, are the pragmatic path here because they make quant selection and model swapping straightforward.