LemonSeed: Qwen3.8-27B Hits 159 tok/s on iPad via eGPU R9700

LemonSeed Studio is an iPad editor that runs on-device inference via an AMD GPU in a Thunderbolt enclosure, embedding the upstream Linux amdgpu driver. Running Qwen3.8-27B Q4 with a DFlash2 draft, the R9700 decodes at 158.5 tok/s on iPadOS and 159.4 tok/s on macOS, versus a 31.2 tok/s baseline, while Strix Halo reaches 64.4 tok/s on Linux.

• DIY Builder Creates Dual RTX 3090 Inference Rig2026-09-26
• 4x AMD R9700 local LLM rig for 10-person startup costs $18K2026-10-11
• Developer compares Gemma4-31B, Qwen3.8-27B, and 6.1-Sol on coding tasks2026-10-11
• ThinkingBox benchmark: one-shot agent success rarely repeats2026-10-09
• Trump unveils $2.4B in tech giant AI pledges2026-10-09