Qwen3.8 27B

User reports Qwen3.8-27B Q6 at about 56.5 t/s generation and 1,032 t/s prompt processing on an Intel Arc Pro B70 32GB at 16k context. Setup is llama.cpp with the SYCL backend, q8_0 KV cache and MTP speculation with 4 draft tokens, on a Threadripper 1950X with 128GB DDR4. At 64k context the same setup gives 756 t/s prompt and 52.7 t/s generation, with a maximum of 223k context on the card. The user compares against RTX 5090 figures taken from other people's posts, not their own machine: about 98 t/s generation and 2,830 t/s prompt, roughly 2x generation and 3x prefill for 5x the price. A 215k-token prompt took 1.9 hours to first token on ollama/Vulkan and 8.6 minutes on SYCL. Qwen3.6-35B-A3B also fits entirely on the card with 262k context at 1,748 / 74 t/s at 16k.