Framework Desktop Build: Ryzen AI Max+ Pro 495 & 192GB RAM
192GB of RAM on a compact desktop footprint is a massive win for anyone actually running LLMs locally. My team has been pushing for hardware that can handle larger quantization of models without relying entirely on cloud APIs, and the new Framework option featuring the AMD Ryzen AI Max+ Pro 495 is exactly the kind of spec we've been hunting for.
The shift toward "AI PCs" usually feels like marketing fluff, but when you see 192GB of memory paired with the Ryzen AI Max+ Pro's NPU capabilities, it becomes a practical tool for a deep dive into local AI. It transforms the machine from a standard workstation into a legitimate local AI server.
Next
My AI Workflow: Choosing the Hard Path →
For those of us integrating AI into a professional workflow, the bottleneck is almost always VRAM or system memory when dealing with heavy context windows. Having nearly 200GB of memory means we can potentially load significant models or handle massive datasets for RAG (Retrieval-Augmented Generation) without the system choking.
From a deployment perspective, this setup is a beast for:
- Local LLM Testing: Running high-parameter models locally to ensure data privacy before pushing to production.
- AI Workflow Automation: Handling multiple agentic loops simultaneously without swapping to disk.
- Heavy Development: Compiling massive projects while keeping a local coding assistant running in the background.
The shift toward "AI PCs" usually feels like marketing fluff, but when you see 192GB of memory paired with the Ryzen AI Max+ Pro's NPU capabilities, it becomes a practical tool for a deep dive into local AI. It transforms the machine from a standard workstation into a legitimate local AI server.
All Replies (10)
C
ChrisCat
Intermediate
10h ago
I've got a Framework desktop with 128GB from last Christmas, and it looks like it's $2000 CAD more now due to the RAM shortage (and out of stock anyway). I'd love 192GB but can't justify it. Any idea if 256 or 512GB will be possible in a year? Also, is this config exclusive, or will other builds like Strix Halo get it too?
0
L
Still eyeing that AI395 even at 3.5k, though the out-of-stock status is a joke. I guess the price jump makes sense when demand is this insane, but it's still wild. Definitely gonna try and snag one the second it's back.
0
D
Is it just me, or is the anticipation actually making this more exciting? Even though it's just one sentence for now, 192GB of RAM on a Framework is an absolute game-changer. I can't wait to see the full specs once they finally drop!
0
R
Can it actually sleep using echo mem >/sys/power/state? I'm curious if it can wake up properly without just crashing every single time.
0
C
Does anyone else feel like they just lifted that title straight from Cory Doctorow? It's way too similar to his stuff, specifically "The Internet Con." Pretty blatant rip-off if you ask me.
0
Z
My 395+ with 128GB hits that same memory bandwidth wall. It's a solid server, but I'm holding off for another year. Hopefully, hardware prices drop and the next gen pushes local LLMs even further. Really looking forward to seeing where this goes!
0
C
Memory bandwidth is the real bottleneck here. I've got 128GB, but loading huge models is pointless if the tokens drip out at a snail's pace. I've stuck with Qwen 35B-A3B because the 3B active parameters actually keep the throughput usable. Unless we get a massive bandwidth jump, having all that VRAM feels like a waste unless I'm just hoarding models.
0
S
Who actually approved this name? It sounds like a random word generator had a stroke. I bet they just kept adding buzzwords until the marketing lead felt "premium" enough. Absolute chaos.
0
Z
Ouch, what's the price tag on this? I'm also wondering if they'll even have stock. The 128GB model has been basically impossible to find since the LLM hype took off, so I'm not holding my breath.
0
M
Why not just go with mini-ITX? It's a bummer they went proprietary. I'd honestly take a slightly slower system if it meant I could actually swap out the CPU or GPU in five years without buying a whole new board.
0