Strix Halo 8060S ROCm 10.0 Qwen3.8-Flash-Next UD-IQ4_XS 64k-depth decode

Protocol deterministic-prime-extend-completions-v1, runner local-llm-speed v1, measured at the OpenAI-completions API endpoint. Mean of 2 repeats after 1 warmup. Concurrency 1, so the figure is per-request, not aggregate. Depth is the primed KV prefix; at depth 0 the whole prompt is computed (actual cached depth 0, recomputed history 0). Prefill is effective follow-up prompt processing: uncached tokens divided by client-observed time to first output. Decode is client-observed generation throughput over the tokens after the first, so TTFT is excluded. KV cache type default, speculative decoding off. Source: local-llm-benchmarks.dev/data/results.json. Run run-08fd1c87a980fba4, completed 2026-09-04. Engine llama.cpp 10.0, runtime ROCm 10.0. Container llama-rocm-10.0. Weights: unsloth/Qwen3.8-Flash-Next-GGUF file Qwen3.8-Flash-Next-UD-IQ4_XS-00001-of-00003.gguf, quant UD-IQ4_XS. Prompt 2048 tokens, 128 generated, primed depth 65536.

Primed depth 65536 plus a 2048-token follow-up prompt. Decode falls from 20.58 tok/s at depth 0 to this figure.