I have 40+ complete runs of DeepSWE from different Qwen3.8 27B models, Flash Next, and MiMo V2.6.
Next, I plan to use one of these models to aggregate the trajectories to produce better results.
I saw several papers at COLM about this. It seems to work.
I'll do the same with Terminal Bench 4.0. I have completed 15 runs so far.