Hello, what would be the best model to run with 48gb vram. Qwen3.6 35B A3B is substantially worse than even cheapest openai models.

I have had a lot of success with qwen3.8-27b UD-Q4_K_XL running 250k context taking up a total of ~37GB reported vRAM usage in nvidia-smi

Yeah, I think 3.8 27B may be the best bet now

I'm using Qwen/Qwen3.8-27B-FP8 on medium reasoning effort and I'm very happy with that. The 35B is a MoE with only 3B active parameters. The 27B is a dense model, so it's slower, but smarter with its 27B active parameters.