Today, we
opened pre-orders for the new 192GB configuration of Framework Desktop with AMD’s new Ryzen AI Max PRO 495+ processor, enabling much larger and more capable AI models to run locally, entirely under your control. This config has both slightly higher clock speeds on the CPU and GPU and runs the LPDDR5X memory at slightly faster 8,533MT/s speeds, both of which enable an increase in AI inference speed over the original Framework Desktop. You can pick up a
DIY Edition, a
pre-built version with Fedora pre-loaded, or the
Mainboard on its own, with shipments starting in November. We’re opening just one batch to start using the memory inventory we have on hand. Pricing and timing for further batches will depend on memory cost and supply updates from our suppliers.
The open-weights model landscape is evolving and advancing rapidly. Today, 192GB lets you run models like DeepSeek-V4.1-Flash Q2 and MiMo V2.6 Flash RL MXFP4. We’re collaborating with Donato Capitella, a software engineer and open source AI content creator on guides and videos for how to get the best results and performance out of your Framework Desktop. He published a
video and
blog post today specifically on models and inference engines to use with the new 192GB Framework Desktop.
Digging deeper into the specifications, Ryzen AI Max PRO 495+ has 16 Zen 5 CPU cores supporting multithreading, 40-core Radeon 8065S Graphics, and a 256-bit memory bus that enables 273GB/s memory bandwidth. In the Framework Desktop, we interface that to two M.2 SSD slots, a x4 PCIe slot, two USB-C ports that support USB4, two USB-A ports, two front Expansion Card slots, two DisplayPort 2.1 ports, an HDMI 2.1 port, a 3.5mm audio jack, 5Gbps Ethernet, and an RZ717 Wi-Fi 7 module.
We changed the PCIe x4 slot to be open-ended on this config, which allows you to use physically larger x8 or x16 cards directly without a riser. That lets you use cards like an Intel E810 for RoCE (RDMA over Ethernet) for 50Gbps networking at ultra low latency, enabling tensor parallelism for linear speedup in inference as you add nodes. As before, the Mainboard is a standard Mini-ITX form factor, meaning you can install it into an existing Framework Desktop case, a
RackMate half-width rack tray, or any ATX-compatible case.
We’ve also introduced a number of firmware improvements with this launch. First, we’ve brought in an BIOS patch from AMD that increases USB4 point-to-point networking throughput from 9Gbps to around 30Gbps. This makes USB4 another great way to cluster together multiple Framework Desktops. Next, we created a much simpler user experience for disabling Secure Boot for those of you who plan to use Arch Linux, CachyOS, and other distros that don’t support it. We’ll be bringing these to all Framework Desktop models soon, and we’re porting the Secure Boot improvements to Framework Laptops too.
Both the DIY Edition and pre-built configurations of the 192GB Framework Desktop come with a 120mm Noctua NF-A12x25 HS-PWM fan pre-installed. Like Framework Laptop 12 (Intel Core Series 3) that we recently launched, the pre-built configuration uses Fedora, making it easy to get started out of the box. This configuration comes with a 2TB SSD, giving you space to download and run large models. Note that the Framework Desktop with Ryzen AI Max PRO 495+ is designed specifically for Linux, so neither the DIY Edition nor the pre-built versions have Windows options available.
With the current memory environment, the 192GB system is a step up in cost over the 128GB Framework Desktop. We expect memory pricing for every capacity to continue to increase over the next six months, so our recommendation if you want to get into local AI inference is to pick up the memory capacity now that matches the models and workloads that you want to run. There are excellent models available across each capacity configuration we offer: 32GB, 64GB, 128GB, and now 192GB. Owning your compute is more important than ever. We’ll continue to do everything we can to empower you on this, no matter what is happening in the rest of the industry.
