Single RTX 3090 Ti achieves high-speed inference with DeepSeek and Qwen models

A user reports running DeepSeek-V4.1-Flash-1017 and Qwen3.8-Flash-Next on six-year-old hardware (RTX 3090 Ti, 128GB RAM), achieving about 35 tok/s and 138 tok/s decode speeds respectively, with an open-source tool unlocking GPU throttles.

• User runs 748B DeepSeek-V4.1-Flash on a €2k PC2026-10-22
• China proposes export controls on 23 AI technologies2026-10-16
• Company Serves 40 People with Local AI on 4x 6000 Pro GPUs2026-10-14
• Reflection AI to release US open-weight model to rival DeepSeek and Qwen2026-10-05
• From one 3090 to 20 DGX Sparks for local LLMs2026-10-04