So Framework Computers had a new reveal event on Feb 25th 2025, and there were a few new developments that are worth mentioning. In case this is the first time you hear about Framework, it’s a young American company that entered the hardware business to make laptops a few years ago, with the brand being all about selling cool, user-serviceable laptops with modular design. They have been fairly successful so far at making a dent on the market (probably still somewhat small compared to companies like Lenovo, Dell and HP). And their support of Linux has been good and fairly serious in the past few years, after focusing first on Windows when they started up. I have already reviewed the Framework Intel 12th Gen laptop 2 years back and the Framework AMD 13 laptop last year, and they keep improving in almost every aspect. So this is with some certain interest that I was looking forward to what they were going to announce next.

Upgrades and Desktop

There is quite a lot to unpack, but let’s make it short because I am sure you guys are busy too.

In a nutshell, they are doing the following:

• Refreshing the line-up with a new top of the line AMD processor for the AMD 13 model.
• Introducing a new desktop system, the Framework Desktop, based on a powerful AMD APU and equipped up to 128 GB of RAM (with options at 32, 64 and 128 for RAM).
• Releasing 12 inches notebooks as well (with touchscreen and fully reversible design a la Lenovo Yoga), aimed at high school students, with different color schemes.

For me the most interesting part was the desktop system, so let’s get right to that.

The Framework Desktop

The Framework desktop is in a very small form factor, using a mini-ITX sized motherboard.

Apparently the system is very silent since it relies on a large fan that makes very little noise (large fans are GOOD, while they take space).

They started talking about the fact that it can be used for gaming, but I can’t really see anyone buying the 128 GB version for that (while the 32GB or 64GB make more sense in that regard, price-wise). The design of the unit is quite peculiar to say the least. In the most mundane way, it’s like a shoebox, but they could not figure out how to make it look cool, and decided to offer these weird, cheap-looking plastic chips that you can re-arrange on your own to customize the front of the desktop. You may have a different taste, but I don’t think it looks good, and it’s somewhat of a downgrade versus the Apple-like design of their laptops.

Apparently the performance of the Framework Desktop for gaming is somewhat similar to what you would get on a RTX4060 Mobile processor, which is good enough for gaming in 1080p, and potentially up to 1440p, but probably not very good for 4k.

It has the following ports, fairly standard, with two USB4 ports and the rest is what you would expect:

And you can order a DIY build if you want to have fun assembling everything yourself (which is usually what I like):

However, the thing that is potentially interesting is the potential for AI (i.e. LLMs for the most part) applications. There are tons of great open weights models nowadays, and sadly there’s a lack of hardware to run them efficiently at home. The problem is like this:

• you need a lot of RAM, if possible VRAM
• you need a lot of processing power, so ideally your model is directly on the GPU
• If you can’t store all your model weights on the GPU, the rest of the model inference will be on the CPU and it will make the performance drop miserably.

But the largest consumer GPU has 32 GB of VRAM (the RTX5090), and cannot run very large models in such a limited capacity, even with quantization. Your option is then to build a multi-GPU build, which is fairly costly and requires a number of specific parts to make it work (many PCIe lanes, proper ventilation, and a big PSU).

The other option is to go for a Mac M-series with a large amount of RAM, since they are build with a very good architecture to accelerate data transfer between the CPU and the onboard memory. But such hardware is expensive, too.

And then you have the DIGITS from Nvidia, rumored to cost around 3000 USD, which should be equiped with 128 GB of RAM/VRAM and use some kind of Nvidia GPU on top of ARM cores. Not sure when it will come out and how many units will actually be available. And it’s aimed at researchers so it could be that you can’t even buy one as a personal user.

And… there’s the Framework Desktop. With 128 GB of RAM, you can fit fairly big models in it. The only question is the memory transfer speed (250 GB/s) and the processing performance. I’d really like to see benchmarks, but what you will end up getting is probably single digits inference on 70b models. That would be my guess. Half-decent for home use, but nowhere near the speed you’d like to see ideally. But the good thing is that… it’s kind of affordable. And they know it, which is why they hammered that message on stage. It was a good call.

In this category, it may be the cheaper option out there to run fairly large models, as long as you understand it’s a trade-off in terms of inference speed.

You can also pair it with other Framework Desktop mainboards with the USB4 connector, while that remains to be seen how useful that kind of setup would be for AI inference. The USB4 transfer speeds will not really help.

The Year of the Linux Gaming Desktop

At the very end they hit the nail on the head by clearly mentioning they believe that 2025 would be the year of the Linux Gaming Desktop. I’m not even making this up!

You can also see two icons on that slide, Bazzite and Playtron (!) which are very good options to transform your PC into a gaming machine, leveraging immutable Linux distributions. It’s kind of amazing to see these two projects showcased in a professional event. Playtron is still in alpha stages, but they are supposed to make some new announcements sometimes this year. And Bazzite is fairly mature solution already working well for handhelds that are not supported yet by SteamOS.

If you want to catch the full event, the video is on Youtube.

Good times ahead!