Spent the last week getting Qwen3.8-27B (bf16, no quant) running on Kaggle's free TPU v5e-8. It works, and it's very fast: ~130 tok/s single stream (MTP), ~78 without 10,300 tok/s prefill, a 105k-token prompt takes ~10 seconds Full native 262,144 context. A 225k-token prompt prefills in ~28s ~540 to
