The most powerful Framework Desktop yet is coming soon with an AMD Ryzen™ AI Max+ PRO 495 processor and 192GB of LPDDR5X memory.
More information coming soon. Join the Framework Newsletter to learn more.
Did you also upgrade the nic to the 10gbit version of the one in the og desktop?
It was not upgraded, however there is the WisePi 10 Gigabit ethernet card (coming soon).
1 Like
“Seize the means of computation” is an EPIC tagline
3 Likes
so since the 395/128gb desktop is out of stock. my guess is both this 495 and 395 wont be on the shelves for at least next 3 months?
192GB of LPDDR5X is how much? US$2000 just for the RAM?
3 Likes
You’ll have to amend your ecom software to charge a price in kidneys 
4 Likes
I wonder where that RAM came from? here?
1 Like
I assume it’s a different type. In that thread I expect it’s upgradeable LPCAMM, whereas for the Desktop, it’s soldered in.
(At a guess, maybe soldered RAM is under less competitive pressure).
3 Likes
They could still use the same chips though, only the form factor is different
Slightly disappointing, with how expensive that thing is gonna be swapping the RTL8126 with a RTL8127 should be a drop-in replacement. Are they still hard to get?
Can I get that in my FW12? 
1 Like
That is quite disappointing for the price point.
GGGgggaaaaaahhhhh there goes my budget for FY2026/2027…
Looking at the current Lpcamm2 price, it would $5000 of RAM
2 Likes
7% faster memory bandwidth? Sounds like its inference speed will be similar to the existing Desktop.
Nice. The 495 is a nice but minor upgrade. It’s mostly a bug fix and correction update to make it work better the 595 is the one that’s gonna be super interesting as that’s gonna be the real next generation, not just minor improvements and fixing problems with the old one.
Still would love to see an AI Max+ 388 version (488 if we’re doing the AMD naming convention?) with maybe 48GB RAM? Would seem like a great “Steam Machine Pro” device.
I’d like to encourage the Framework team to consider adopting 10GbE instead of 5GbE for the Ethernet interface.
AMD’s official Strix Halo documentation recommends using 10GbE networking when connecting multiple systems together to run larger LLM workloads. While I’m not certain how much of a bottleneck the current 5GbE interface on the Framework Desktop would be in practice, matching AMD’s recommended networking specification would improve compatibility with this use case and provide greater flexibility for users building multi-node AI systems.
The bottleneck in general for distributed inference is latency rather than throughput, so regardless of 10GbE or 5GbE, without support for RoCE, latency will be too high to support tensor parallelism at high performance. Part of why we have a PCIe slot on the Mainboard is to enable connecting NICs that support RoCE, like an Intel E810.
4 Likes