Intel B65 hits 62 tok/s on Qwen3.8-Flash-Next with XeStrata 4-bit offloading

SamSausages · reddit · 2026-10-31

A Reddit user benchmarked a single 32GB Intel B65 running Qwen3.8-Flash-Next with XeStrata Flash-Next 4-bit, offloading experts to system RAM:

• Single user: 62.4 tok/s median and 2.1s TTFT with tuned settings (PCIEFRAC 0.5, 65k context); 57.5 tok/s with recipe defaults
• Concurrency: aggregate throughput drops to 48.7 tok/s at 2 users and 44.6 tok/s at 4 users, with TTFT rising to 5.2s/11.4s — best suited to a single user
• Comparisons: same weights on an RTX A6000 with llama.cpp MTP-2 hit 52.8 tok/s single-user but 75.2 tok/s aggregate at 2 users, clearly better for concurrency. B65 via llama.cpp SYCL only managed 17.9 tok/s, making XeStrata a huge win
• The author argues Intel's entry-level lineup has improved dramatically in 6 months and, despite a recent price bump, remains the best entry-level local inference stack with enough system RAM

Full interactive charts and raw data are on the author's site.

Original post →

More from Infra

• Anthropic S-1 math: subscriptions are ~18% of revenue but burn over half the compute — teortaxesTex · 2026-10-31
• Reddit: 4-bit quantization could cut memory 75%, but incentives block it — Excellent_Place4977 · 2026-10-31
• NVIDIA discloses Brev platform breach: SSH keys exposed Oct 4–20 — HowDevelop · 2026-10-31
• Northflank hacked: users told to rotate all secrets as AI keys confirmed abused — basedjensen · 2026-10-31
• Arm exec: AI datacenters need a third 'agentic CPU' class as DPUs become security table stakes — BenBajarin · 2026-10-31
• Seroter's 900th reading list: hill-climbing agents and prefix caching costs — rseroter · 2026-10-31

Infra channel →