GMKtec AI Mini PC AMD Ryzen Al Max+ 395 (up to 5.1GHz) Mini Gaming Computers, 96GB LPDDR5X 8000MHz (8GB*8) 1TB PCIe 4.0 SSD, Quad Screen 8K Display, WiFi 7 & USB4, EVO-X2

Powerful little beastie, truly amazed at the benchmarks on it. Cinebench was better on this than our 4 year old threadripper, gpu benched between a 4070 and 4070 ti. With the price of memory on desktop systems right now, this is priced like buying the memory and getting a powerful cpu and a decent gpu and it just sips power. The camera server that this is replacing would only run for 2 hours on our ups before, estimated time on our ups now is 2 days. This minipc also uses a standard pc fan instead of proprietary cooling, making replacement a breeze (if anyone who ever worked with minipc's knows, it's usually the cooling that goes first). I do wish it came with basic peripherals.

Using Openclaw and LM Studio 0.4.9 CPU/GPU: AMD Ryzen AI Max 395 (Strix Halo/Radeon 8060S, 40 CUs) Sweet Spot: Why Qwen3.5-35B-A3B---- MoE structure means only ~3B parameters active at any time despite 35B total. This feels like a smaller model with deep reasoning capability. Perfect balance between intelligence and speed - anything more would just take forever! Optimal LM Studio Inference Settings Core Parameters (Already Optimized) NGl (GPU Layers): 99/All ✅ Max GPU offloading Quantization: Q4_K_XL ✅ Best balance of speed/quality Context Length: 64k default, 122k for deep reasoning Batch Size: Default or slightly higher Sampling Parameters Temperature: 0.7 standard (0.3 reasoning, 1.2 creative) Top P: 0.9 (0.8 tighter, 0.95 looser) Top K: 40 (30 reasoning, 50 creative) Repeat Penalty: 1.1 standard Advanced Features Min P Sampling: 0.05 ✅ Reduces hallucination risk Speculative Decoding: On with draft model <7B ✅ Speed boost ~2x on repetitive patterns Draft Cutoff: 0.15 ✅ Balanced speed/quality Max Draft Size: 12 tokens ✅ Moderate speculation 💰 Price Analysis () Was it worth it? YES for my use case (local LLMs): Specialized capability: massive models on integrated GPU 96GB unified memory handles large contexts without swapping Not generic fast - but uniquely capable for LLM work The 128GB variant isn't worth extra cost, as CURRENTLY with (Qwen3.5-35B-A3B) you really wouldn't want to run a model more then this anyways.