Someone finally spent ~$8,000 on the new 192GB Ryzen AI Max+ PRO 495 Minisforum and tested the crap out of it. He spent the weekend running DeepSeek, MiniMax, GPT-OSS and Qwen on it. Now he's boxing it back up and returning it. 👀 Interestingly, the machine works as intended. With llama.cpp/Vulkan 🧠 GPT-OSS 120B → ~58 tps 🧠 Qwen3.8 Flash-Next → ~28 tps 🧠 MiniMax M2.7 230B-A10B → ~31 tps 🧠 DeepSeek V4 Flash 284B-A13B → ~18 tps That 192GB unified memory does exactly what you want 💾 Qwen3.8-27B @ 262K → 33.3GiB 💾 GPT-OSS 120B @ 131K → 61.3GiB 💾 DeepSeek V4 Flash @ 131K → 108.9GiB 💾 MiniMax M2.7 @ 196K → 127.4GiB He also found some huge optimization wins. Qwen3.8-27B: 🐌 no MTP → 9 tps 🚀 MTP → 24 tps That's 2.6× faster from speculative decoding. And cached agent turns changed the experience completely: GPT-OSS → 35 sec → 1.9 sec Qwen → 106.6 sec → 1.2 sec MiniMax → 131 sec → 4.4 sec DeepSeek → 164 sec → 1.6 sec So why return it? Ryzen AI Max+ 395 MINISFORUM 💰 $3,879 💾 128GB 🚀 256GB/s Ryzen AI Max+ PRO 495 💰 $7,399 💾 192GB 🚀 273GB/s That's roughly 91% more money for 50% more memory and only ~7% more memory bandwidth. The extra memory unlocks models the 128GB box can't comfortably hold. It doesn't make inference twice as fast. His final verdict was still “I am boxing this back up and sending it back. Not worth $8k!” 🔗 Original owner review in ALT.