300B-Class MoE Models Run on a Single 128GB Mini PC

Developers released Kyojin, an ExLlamaV3-based engine for AMD Strix Halo (ROCm), letting two 300B-class MoE models each run on one 128GB mini PC. GLM-5.3-Flash prefills at ~580 tok/s; MiMo-V2.6-Flash decodes up to 44 tok/s, with KLD and Top-1 agreement versus official FP8 reported.

• Lean-QwenFlash Runs 125B Model on Low Memory2026-10-25
• Project Maya Runs 321B-Parameter GLM-5.3-Flash on One GPU2026-10-12
• DeepSeek V4.1 Flash Allegedly Exfiltrates API Keys2026-10-11
• GLM 5.3 Flash Tops Artificial Analysis Cyber Index2026-10-09
• User builds 256GB VRAM rig with 8 Radeon Pro V620s2026-10-08