Struggled a bit with the dedicated BIOS of 96gig vram or setting vram low and just having system ram say 110gigs but this was mostly cos of LM Studio, I found running Lemonade AI server is much better suited and built for this device. You then run the bios setting of 96gig VRAM and inside Lemonade AI server it actually filters models when u search to be compatible with this device. Lemonade really works well man. Top notch AI server.

You can pass in context size, all sorts of llama cpp args etc. I'm running "unsloth/Qwen3.5-35B-A3B-GGUF:Q8_0" and it is working brilliantly. Around 25-30 tok/s. Also been running "lmstudio-community/Qwen3-Coder-Next-GGUF:Q6_K" - Both these models work brilliantly. The Coder Next model doesn't play nice with tools like GSD tho.

Been running them all in Opencode and it works perfectly for agentic coding. I also run ": ggml-org/gpt-oss-120b-GGUF:MXFP4" for chatting via the lemonade AI chat interface for asking debugging question. All in all a serious piece of kit this.

As per the title, it is fast and powerful but large for a NUC. Runs AI well, and once replaced, it'll still make for an awesome Steam Box analogue. This model has some life ahead of it.

Wanted to try out larger LLM models that would not fit into my 4080s 16GB Vram and not being an Apple fan (although thier unified memory is tempting) I figured I would try this 395+ chip and 128gb ram and see how well it works. Simple answer is that I have loaded models that needed 30gb+ Vram and they all load and run successfully. Granted, the LLMS dont reach the speeds of multiple 5090s with 32gb vram tokens per second, but they run which is what I wanted... my 4080 system would just crash and burn. Other than that use case, the reviews praising the 395+ and its 8060s gpu are confirmed... Its snappy doing office stuff and 1080p gaming is totally acceptable at decent quality settings. Now wish I could afford a bunch more for a totally overpowered proxmox HA cluster but that would be stupid..

Got it specifically to run some larger LLMs locally without breaking the bank and it does exactly that. Its not as fast as a decent GPU with enough VRAM obviously, but unlike the GPUs with limited VRAM it CAN actually load 32b+ LLMs and they run without the system crashing... and at 20t/s its ok and usable for testing. As a system for more general use, even 1080p gaming, its fantastic... The Asus ROG Flow Z13 (2025) laptop must be a beast with this chip and memory config.