Two 96GB Ascend Cards Run Qwen3.8 Flash-Next

A user built a local inference machine with two Huawei Atlas 300I Duo cards (96GB each, actually two 48GB chips per card). After vLLM, driver and operator optimizations, Qwen3.8 Flash-Next went from about 1 token/s to roughly 30 tok/s single-request and 61 tok/s at four-way concurrency, completing all 198 GPQA Diamond questions.

• User builds 256GB VRAM rig with 8 Radeon Pro V620s2026-10-08
• Offload most of Qwen3.8 KV cache to RAM2026-09-16
• Qwen3.8-Flash-Next gains Strix Halo support on Strata2026-10-07
• Qualcomm Licenses Huawei LogicFolding Chip Tech Patents2026-10-05
• Developer releases Ninfer 4080 for 16GB GPUs with 100k context2026-10-04