Comment by SillyUsername

Maybe they should have tried Qwen 3.8 Flash Next and Strata. Faster than 27B, around 60tps, and more intelligent on benchmarks.

I'm using this on my own 2x 5060TI.

I abandoned 27B on llama - split mode tensor, 200K context (Q5/Q4_1), ubatch 3072, batch 6144, it managed 1900tps prompt processing (same result as article for decode).