User runs 748B DeepSeek-V4.1-Flash on a €2k PC

A user ran the 748B-parameter DeepSeek-V4.1-Flash-1017 at 7.2 tok/s on a used ~€2,000 PC (RTX 3090, 128GB DDR4, i9-9900X) using a 1-bit quant and llama.cpp. A 110GB Engram table stays on NVMe and is read on demand, letting it fit. The output contained a physics error and has little practical use, the user said; Qwen models work better on the setup.

• Community tests full n-gram table for Qwen 3.8 Flash Next2026-10-18
• From one 3090 to 20 DGX Sparks for local LLMs2026-10-04
• Yandex Releases 80B AliceAI Foundation Model2026-09-22
• SCMP: Huawei, DeepSeek narrow chip ecosystem gap with Nvidia2026-10-22
• DeepSeek open-sources its DeepInfer inference engine2026-10-20