Qwen3.8 27B

User reports Qwen3.8-27B at 37.8 tok/s on a Ryzen AI MAX+ 392 laptop with Radeon 8060S and 32 GB, against 29.6 tok/s for upstream gufo on the same machine. Setup is the user's tubo fork of gufo with IQ3_XXS weights and speculative decoding; the 512-token answer figure is a 10-run average. The apex quant with turbo6 6-bit KV cache reaches 45.1 tok/s on the same test and 53.3 tok/s over the HumanEval run, with 17.6 GiB RAM at 4k context and 128k context tested on 32 GB. Prefill is no faster than gufo up to about 1k tokens; splitting work onto the NPU gives 484 vs 440 tok/s at 4k and 472 vs 428 tok/s at 8k prompt tokens.