Hardly any benchmarks for Intel B65, so I got 4x and tested them. Long Context Sweeps for Flash-Next and other models

EDIT: Argh, forgive the formatting, .md and reddit are getting the better of me today! Howdy Folks! New batch of tests from my 4x B65 rig I built, because I was curious to test how the high memory bandwidth works out in reality. This time doing long context sweeps for our most popular models! If there are specific tests that you want to see, hit me up and I'll see if I can fit them in. I'm also starting offload tests, where I put some of the layers into system RAM. So far only doing it with flash-next, but interesting results in the "by model" section of my site for flash-next. Some Examples below, full context sweep list of all the models I tested: b65.thekaiser.us/matrices AI Generated table from my data: Qwen3.8 context sweeps on Intel Arc Pro B65 Each cell is total throughput / median per-request throughput , in tokens/sec. C1–C16 = requested concurrency. Context lengths use K = 1,024 tokens. = requests queued (capacity-limited). — = skipped as over capacity, not zero throughput. Qwen3.8-Flash-Next-W4A16 — 4× B65 — INT4 vLLM 0.30.0 · October 18, 2026, 01:06 UTC run Context C1 C2 C4 0 84.5 / 87.7 145.6 / 76.4 243.2 / 63.0 16K 73.4 / 84.7 115.4 / 67.0 166.4 / 45.9 32K 71.0 / 91.6 105.7 / 66.6 137.3 / 41.0 64K 68.6 / 78.7 109.3 / 66.3 143.7 / 41.5 128K 55.0 / 82.1 72.3 / 49.5 — 200K 45.8 / 71.6 66.2 / 49.3 — Full results Qwen3.8-Flash-Next-W4A16 — 4× B65 — INT4 vLLM 0.30.0 · October 18, 2026, 00:31 UTC run Context C1 C2 C4 0 42.7 / 43.4 85.1 / 43.6 159.2 / 41.2 16K 40.8 / 42.3 79.4 / 41.9 141.9 / 37.8 32K 39.3 / 41.6 75.2 / 41.3 127.3 / 34.9 64K 38.6 / 40.4 73.5 / 38.9 127.6 / 34.5 128K 34.5 / 38.1 61.2 / 35.2 101.7 / 28.9 200K 31.5 / 35.9 57.0 / 32.3 — Full results Qwen3.8-27B — 1× B65 — INT4 vLLM 0.31.0+xpu.llmscaler · October 19, 2026 Context C1 C2 C4 C8 C16 0 20.2 / 20.4 59.9 / 31.2 115.1 / 30.0 204.3 / 27.5 159.3 / 10.4 8K 19.8 / 20.0 57.4 / 29.5 106.6 / 28.1 184.5 / 24.5 144.8 / 9.6 32K 18.2 / 18.8 48.2 / 25.9 84.0 / 22.7 — — 64K 16.7 / 17.4 42.1 / 22.4 — — — 128K 14.6 / 15.2 — — — — Full results Qwen3.8-27B — 4× B65 — INT4 vLLM 0.31.0+xpu.llmscaler · October 19, 2026 Context C1 C2 C4 C8 C16 0 50.4 / 50.9 129.8 / 66.8 253.4 / 66.1 468.2 / 62.1 458.1 / 30.1 8K 49.4 / 50.4 125.9 / 65.1 239.6 / 63.6 417.1 / 56.6 422.8 / 28.2 32K 46.2 / 48.5 108.7 / 58.9 197.7 / 56.9 310.6 / 45.3 332.8 / 24.1 64K 43.5 / 46.3 96.3 / 53.3 172.2 / 47.8 270.9 / 39.0 286.2 / 18.9 128K 37.8 / 42.5 76.0 / 44.8 121.8 / 38.5 181.8 / 26.9 — 200K 33.8 / 38.9 63.3 / 38.2 97.9 / 29.1 — — Full results Qwen3.8-27B — 4× B65 — FP8 vLLM 0.32.0 · October 19, 2026 Context C1 C2 C4 C8 C16 0 30.5 / 30.8 59.7 / 30.4 116.2 / 30.0 222.8 / 29.2 404.2 / 27.7 8K 30.0 / 30.4 58.6 / 29.9 111.6 / 28.6 205.8 / 26.4 354.5 / 23.1 32K 28.7 / 29.7 54.4 / 28.4 98.5 / 26.1 177.4 / 22.4 269.7 / 17.6 64K 27.9 / 28.9 52.3 / 27.2 94.0 / 24.9 164.3 / 21.3 242.3 / 16.2 128K 25.3 / 27.3 46.3 / 24.6 78.9 / 21.4 117.2 / 16.7 — 200K 23.2 / 25.7 41.3 / 22.6 66.9 / 18.7 — — Full results Qwen3.8-27B — 4× B65 — BF16 vLLM 0.32.0 · October 19, 2026 Context C1 C2 C4 C8 C16 0 28.6 / 28.7 56.2 / 28.4 110.0 / 28.0 211.0 / 27.2 389.5 / 25.6 8K 28.1 / 28.4 55.2 / 28.0 106.2 / 27.1 197.9 / 25.3 357.6 / 22.5 32K 27.1 / 27.7 52.1 / 26.9 96.5 / 25.3 176.9 / 22.4 287.0 / 18.3 64K 26.0 / 27.0 49.5 / 25.9 91.3 / 24.1 159.1 / 20.7 — 128K 24.0 / 25.7 45.1 / 23.8 79.3 / 21.3 — — 200K 22.7 / 24.3 40.3 / 21.9 67.7 / 18.8* — — Full results These are sustained-decode measurements from 30-second tests. Per-request values are the recorded medians, not total throughput divided by concurrency.