Running Qwen3.6 35B A3B for coding with 20GB of VRAM
A laptop with 64GB of DDR5 memory and a Radeon RX 7900 XT external with 20GB of VRAM was used to run the pi coding agent locally. worked with higher-precision quantization by combining its design with . The Q8 version fit with a 100,000-token , while UD-Q6_KL supported the full 262,000-token window.
It was slow but produced more than 20 , which was considered practical. It remains unclear whether another model would work better or whether a lower-precision dense model placed entirely in VRAM would offer a better balance.
Key points
- The setup has 64GB of and 20GB of VRAM.
- ran at Q8 with a 100,000-token .
- UD-Q6_KL allowed the full 262,000-token .
- moved part of the workload outside the .
- Generation above 20 was practical for local use.