Expert’s Rating
Pros
- Extremely detailed meeting directions
- Simple however configurable
- Powerful AMD Ryzen AI Max+ chip inside
- An total sense of high quality
Cons
- Memory is soldered and isn’t upgradable
- Attaching the SSD takes some nerve
Our Verdict
The Framework Desktop DIY Edition is a thoughtfully engineered small-form-factor desktop PC that’s each an entry level into fanatic computing in addition to a strong AI desktop in its personal proper.
Price When Reviewed
This worth will present the geolocated pricing textual content for product undefined
Best Pricing Today
Best Prices Today: Framework Desktop
The Framework Desktop DIY Edition is exclusive: a do-it-yourself desktop with out the complexity of constructing from scratch, forming a compact, customized “AI workstation.” If you’re nervous a few less-familiar model, don’t be. Framework obsesses over particulars, from manuals to useful hints etched into the supplies themselves. Multiple photographs present tips on how to tighten a thumbscrew–that’s how comfy they need you to be.
I can level to a couple issues that I assumed wanted enchancment: soldered reminiscence, a beta driver bundle that must be finalized by the point you purchase it, and a high panel which didn’t clip in as simply as I’d have favored. Inserting the SSD harassed me out a bit, too.
But Framework’s eye for personalisation (coloured tiles you may design and set up your self, plus your selection of I/O) lends itself to enjoyable and productiveness. The AMD Ryzen AI Max+ (Strix Halo) chip inside is barely out of the abnormal, with its do-everything design. I’ve excessive reward for the Framework Desktop, and suppose you’ll too. Let me let you know why.

Mark Hachman / Foundry
Framework Desktop: Specs
Framework lets you pick and choose which elements you wish to add… form of. Our evaluate mannequin was a DIY Edition, which asks you to construct the PC your self. Framework additionally sells a completely pre-built mannequin, too.
For every Desktop, you’ll want to pick out which Ryzen AI Max processor (with an built-in GPU) you like: the Max 385, an 8-core chip with 32GB of LPDDR5x reminiscence connected, as much as the Max+ 395, a 16-core processor with a whopping 128GB of reminiscence. Each processor is pre-mounted on the Framework mainboard, contained in the Desktop’s case and with an influence provide. It’s all upgradable, with one exception: the reminiscence is soldered on to the motherboard.
Because of this, costs start at $1,099 for AMD’s Ryzen AI Max 385 and climb to $1,999 for the Max+ 395 possibility. While this DIY Edition permits you to cease and contribute your individual storage, you most likely received’t.
So begin tacking on the extras: WD Black SSDs bought by Framework (between $69 and $699 for 500GB to 8TB) function main and optionally available secondary storage, as much as $199 for a Windows 11 Pro license, and choices like a translucent facet panel, customized tiles, and so forth add value. You’ll want to choose considered one of three 120mm CPU followers, too, or provide your individual. Our evaluate mannequin (detailed beneath) value somewhat over $2,500.
To recap: assuming you provide your individual SSD and working system, you may spend as little as $1,099 for the Framework Desktop. But you’ll most likely wish to purchase some storage from Framework, too. Just don’t overlook the optionally available growth playing cards ($11-$39 apiece) and the coloured tiles ($10-$15 for a pack of seven monochrome tiles, and $5 apiece for brand tiles). They’re optionally available however make the Framework enjoyable.

Mark Hachman / Foundry
Framework’s Framework Desktop configuration page spells all of it out. Be conscious, although, that Framework makes the Desktop in batches, and the web page reveals the Framework Desktop presently bought out by someday after October.
(As of press time, Framework had not issued any up to date steering on President Trump’s proposed chip tariffs of 100 %. Check the Framework tariff blog for up to date info.)
- Processor: AMD Ryzen AI Max 385/Max+ 395
- Graphics: Radeon 8060S (built-in)
- NPU: Yes, 50 TOPS
- Memory: 32GB (Max 385)/64GB (Max+ 395)/ 128GB (Max+ 395), LPDDR5X-8000, soldered to motherboard
- Storage: 500GB to 8TB NVMe PCI 4.0 SSD (2TB as examined)
- Power provide: 400W, Gold (110V), or Silver (230V)
- Ports: 2 USB-C (USB 4 40Gbps), 2 DisplayPort 2.1, 2 USB-A 10Gbps, HDMI 2.1, 5Gbit Ethernet, 3.5mm jack
- Security: none built-in
- Wireless: Wi-Fi 7, Bluetooth
- Operating system: User-supplied (Windows 11 Pro as examined)
- Dimensions: 8.9 x 8.1 x 3.9in.
- Weight: 6.8 kilos
- Color: Black
- Price: Starting at $1,099, $2,515 as examined
It’s not only for tinkering — this machine can legitimately run the most recent AI fashions domestically, one thing few desktops this dimension can do.
Framework Desktop: Build course of
Our evaluate unit of the Framework Desktop arrived in what nearly gave the impression to be a “flat pack” field such as you’d discover from Ikea. (Framework requested the way it must be custom-made; I responded with a picture of our PCWorld brand and a request to “make it cool.”)

Mark Hachman / Foundry
Each part got here individually boxed, with the highest field bearing an HTML glyph and hyperlink to the web quick-start build instructions. This course of was extraordinarily simple, insanely detailed, and leaned into the a part of my psyche that appreciates exact, well-documented directions. If you’ve ever struggled by an Ikea handbook, your soul will likely be soothed by the Framework guides, assured.
The Framework Desktop measures 8.9 x 8.1 x 3.9 inches, which interprets into 4.5 liters of area. The Desktop will match inside an abnormal backpack, although not as conveniently as a laptop computer.
As my construct article notes, Framework’s DIY Edition actually doesn’t offer you that a lot to really do, apart from insert the SSD, connect the fan, and mount the facet panel, tiles, and growth playing cards. Normally, “building” a PC entails putting the CPU on the motherboard, making use of thermal paste and a cooler, and inserting/attaching the motherboard to the case. Not right here.
Framework does this all for you already, mounting the mini-ITX Framework Mainboard contained in the steel chassis, together with a FlexATX 400W energy provide with both a Gold (110V) or Silver (230V) score. You should set up the SSD your self; the first M.2 slot is effectively inside the case, making it a bit difficult to put in. A secondary slot is on the underside of the motherboard and simply accessible by way of a facet panel.

Mark Hachman / Foundry
Framework’s specifications page goes into extra element than you would possibly want however will most likely respect, such because the acoustic noise and airflow rankings of the three optionally available CPU followers from Cooler Master and Noctua right down to the scale and design of the heatsink. Speaking of airflow, the Desktop operated quietly with little or no noise. Inside, all the cables have been tightly wrapped and out of the best way.
Framework even goes as far as to publish different guides that you could be by no means use, akin to removing the Framework mainboard from the Framework case and putting in it in a completely new, customized case. If you’re thinking about a extra detailed dialogue of the construct course of, take into account studying my separate article.

Alex Esteves / Foundry
Framework says that the entire course of will take you about 45 minutes, which sounds about proper. Remember, that features putting in an working system like Windows, which may simply take 20 to 30 minutes by itself.
My set up additionally shipped with beta drivers that AMD hadn’t totally signed off on. That not solely meant that I needed to set up a driver bundle (for Wi-Fi, audio, and the like) in addition to the working system, however that they affected the efficiency as effectively. (Framework mentioned the night time earlier than this evaluate printed that it had begun delivery finalized drivers.)
Since AMD’s Ryzen AI Max CPU is a cellular half, the chipset is built-in into it. That definitely helps the Desktop from an upgradability standpoint, as that facet is taken care of.
Most of the facet panels connect to by way of a mixture of magnets and/or tabs, which went on easily, however required a little bit of fiddling. The solely challenge I had was with the highest panel. Framework used some sliding tabs to carry the highest panel down, together with two thumbscrews. For no matter motive, the tabs both refused to slip in for minutes at a time, or did so instantly. I’d encourage Framework to seek out one other answer.

Mark Hachman / Foundry
Don’t overlook to order and set up the coloured tiles! Those are the soul of the Framework desktop, and are available in numerous colours. Each tile (save for a few of the customized, logoed tiles) has a couple of slim slits for airflow. I considered these as a mud display of kinds till I observed that air is sucked in by the grillwork on the facet panel, pushed throughout the CPU utilizing the CPU fan, then pushed out by the back and front.
Many folks like to embellish their laptops utilizing stickers. I hate doing so, however actually had enjoyable utilizing the coloured tiles to create a enjoyable little mosaic highlighting the PCWorld customized tile that Framework commissioned.

Mark Hachman / Foundry
Again, the tiles are optionally available. Airflow is sucked in by the facet grille and pushed out by the back and front, so make certain and reserve a lot of the slatted tiles for the highest of the entrance panel to keep away from blocking the airflow because it exits the case.
Framework requested us for a customized tile design with our PCWorld brand. It doesn’t lengthen customers the identical alternative, at the least but.

Mark Hachman / Foundry
The Desktop’s rear ports embrace an HDMI 2.1 port, a pair of DisplayPort 2.1 ports, two USB4 connections, a 5Gbit Ethernet port, two USB-A ports, and a 3.5mm headphone jack. Thunderbolt ports utilized by Thunderbolt docks are simply branded USB4 ports, so the 2 USB4 ports will hook up with a USB4 or Thunderbolt 4 dock at 40 Gbps speeds.
Framework additionally permits you to purchase a number of I/O connectors, referred to as Expansion Cards, which plug into two entrance receptacles. Each growth card is actually a passive USB-C dongle, and you should purchase USB-A, Ethernet, headphone jacks — all the things apart from a show port. Each growth card slots into the slot, then latches utilizing a small lock on the underside of the case. Everything labored completely by way of efficiency and swapping them in or out.

Mark Hachman / Foundry
Framework Desktop: Performance
Testing the Framework Desktop is a problem, if solely as a result of it may well achieve this many issues effectively. It’s each a productive (and quiet) desktop machine, however its built-in GPU is highly effective sufficient for after-hours gaming. But the actual motive the Framework Desktop could exist is due to its huge quantities of VRAM, which lends itself to downloading and operating numerous iterations of AI, together with AI artwork and chatbots, or LLMs.
AMD’s Ryzen AI 395+ helps a TDP of 55 watts, present as each a cellular processor in addition to a desktop CPU. To date, the one different machines to make use of AMD’s chip have been gaming tablets just like the Asus ROG Flow Z13 or extra conventional clamshell notebooks just like the HP ZBook. Manufacturers have begun to discuss with a few of these as AI workstations, regardless that they don’t actually mirror the massive, cumbersome workstations of outdated. Framework’s Desktop certain doesn’t.
Desktop graphics playing cards ship with devoted video RAM; in laptops, VRAM is allotted a special approach. The incontrovertible fact that the Ryzen AI Max (Strix Halo) is a cellular processor is extra necessary than you would possibly suppose.
In AI, the video RAM connected to a GPU is actually the RAM the AI algorithm makes use of. Laptops with Intel chips inside divvy up the system RAM equally between the GPU and the working system: an Intel Core Ultra pocket book with 32GB of RAM makes use of 16GB to execute packages and assigns 16GB to the GPU as VRAM. AMD does it in another way: the BIOS assigns what VRAM it thinks it wants, however you may manually modify it, one thing you can not do with most Intel-based laptops.
Why is that necessary? Out of the field, the Framework Desktop wasn’t actually optimized to benefit from its huge quantities of RAM. Remember, Framework issued reviewers early drivers. As it turned out, these drivers didn’t ship the most effective efficiency, particularly on AI purposes. AMD (coincidentally or not) printed a beta driver to its Adrenalin software program bundle, which permits the 128GB of RAM connected to the AI Max+ 395 to assign 32GB to the working system, leaving a whopping 96GB as VRAM for graphics or AI.
Basically, what I’ve achieved is first examine the Framework Desktop to quite a lot of present “Strix Halo” PCs, all within the cellular area, to display its comparative efficiency. Next, we’ll examine the Framework Desktop to quite a lot of gaming laptops — not a lot to level out which is best, however to permit you to mentally classify the place the Desktop and its Strix Halo chip sits within the grand scheme of issues. Which gaming laptops does it examine to? I feel that’s as helpful as the rest.
For the primary batch of comparisons, we’re pitting the Framework Desktop towards the HP ZBook Ultra G1a, basically a direct comparability with a Ryzen AI Max+ Pro 395 chip inside and 128GB of RAM; the $2,299 Asus ROG Flow Z13 gaming tablet with a Ryzen AI Max+ 395 however with solely 32GB of RAM; and a model of the $1,699 Asus ZenBook S16, with a a lot lower-performance Ryzen AI 9 HX 370 and 32GB of RAM.

Mark Hachman / Foundry
In sure purposes, we’d anticipate that the higher cooling of the Desktop would possibly make a distinction; in different benchmarks, the extra VRAM would possibly give it an edge. Cinebench is a purely CPU-driven benchmark, which might be run on a single thread — the place efficiency is statistically equal — and throughout all the cores and threads.
Given that three of those units use very comparable CPUs, you would possibly anticipate that these numbers could be just about an identical. But there’s a 19.6 % improve from the ZBook to the Framework Desktop in CPU efficiency alone. In a case like this, I’d attribute the distinction by way of the Desktop’s superior cooling.

Mark Hachman / Foundry
I didn’t have private entry to all of those units; we preserve a opinions database, and sure units are shipped backwards and forwards. In this case, we did have entry to the Blender benchmark, which makes use of the open-source 3D creation instrument to generate three totally different scenes. This check measures each the CPU and GPU.
Again, we’re seeing an identical division between all 4 units. Looking solely on the Monster benchmark (proven in blue), the Desktop improves by 27.6 % over the HP ZBook, with no change to the scores due to the improved driver. That will change.

Mark Hachman / Foundry
This was after I began to ask questions on whether or not Framework’s driver made a distinction. Procyon’s Image Generation check merely creates a sequence of 16 1024×1024 photos, utilizing a standard immediate. Framework’s Desktop wasn’t sluggish to take action, but it surely fell behind a few notebooks, which made me elevate my eyebrows. As it turned out, AMD’s driver made extra RAM accessible to the processor and the picture era algorithm.
Here, the Framework Desktop improves by 26 % over the HP ZBook, roughly constant to the way it behaved within the Blender benchmark. The Framework Desktop generated a picture about each ten seconds, which is fairly good with no discrete GPU. Obviously, you’ll see higher (or worse) efficiency operating AI artwork on different fashions not utilized by this benchmark.

Mark Hachman / Foundry
I didn’t have all the units accessible to me to run this benchmark, which approximates an AI chatbot / LLM, operating 4 comparatively outdated fashions. The check measures the pace of the output, asking comparable questions or prompting comparable duties.
This benchmark was fairly shut utilizing Framework’s driver, however with the improved AMD driver the efficiency bounce on the Phi 3.5 benchmark will increase simply 8.7 %. Generally talking, chatbot efficiency might be going to be fairly comparable throughout these units, even these we didn’t check.

Mark Hachman / Foundry

Mark Hachman / Foundry
We’ll have some extra gaming benchmarks a bit decrease down, however these two video games point out that gaming is definitely doable at excessive ranges on the Framework Desktop. We’re exhibiting right here each common framerates in addition to the “minimum” one % framerates, indicating that if and when your recreation stutters, even a framerate drop will put it at playable ranges. Most avid gamers take into account 60 fps to be absolutely the minimal, with above 90 fps most well-liked.
At over 90 fps for each video games on the highest visible settings, each of those comparatively current video games are definitely playable on this PC.

Mark Hachman / Foundry
For different video games, we use 3DMark’s benchmarks as a proxy for the way effectively the system will carry out.
In this case, developer UL recommends that we use the Steel Nomad Light benchmark as the suitable one for this {hardware}. Here, the Framework Desktop demonstrates a 16.5 % improve versus the HP ZBook.
Next, we in contrast the Framework Desktop to quite a lot of gaming laptops. Again, the aim right here was to not confirm whether or not the Framework Desktop was the most effective in any specific class, however to offer you a way for the place the Desktop and its Strix Halo chip stands by way of {hardware} that you simply already know and perceive.
For this, we in contrast the Desktop to the $1,499 Alienware 16 Aurora, which makes use of a Core 7 240H chip on Intel’s Raptor Lake structure, and a Nvidia GeForce RTX5060; in addition to two Intel “Arrow Lake” notebooks. The first, the $3,299 Asus ROG Strix Scar 16, incorporates a Core Ultra 9 275HX and an RTX 5080. The different, the $4,499 MSI Raider 18 HX A2XW, features a Core Ultra 9 285HX and an RTX 5080 as effectively, and is without doubt one of the quickest notebooks we’ve seen.
Finally, we’ve added the slower $1,199.99 Dell G15 with an Intel Core i7-13650HX chip and an RTX 4060 chip inside, plus the $2,699 Razer Blade 14 with an AMD Ryzen AI 9 365 CPU and an RTX 5070 GPU.
The AMD Ryzen Max+ 395 contains 16 cores and 32 threads, and runs as much as 5.1GHz within the turbo mode. That really places it forward of the Core Ultra 9 contained in the MSI Raider, which has 24 cores and 24 threads. The distinction is that these threads run as much as speeds of 5.5GHz, most likely giving the Raider the sting right here.

Mark Hachman / Foundry
We usually, select PCMark 10 as our “how well will this system perform everyday tasks?” check. It’s a conglomerate of exams starting from video chat to internet searching, workplace work, with capabilities like CAD thrown in for good measure. Most laptops and desktops excel right here, however the extra “heavier” duties like CAD can drag down underperforming methods.
There aren’t any points right here. In common work, the Framework Desktop will likely be one of many quickest machines on the planet.

Mark Hachman / Foundry
Handbrake converts a feature-length movie of about 90 minutes right into a format appropriate for a pill. It’s an outdated benchmark from the standpoint that streaming is widespread on airplanes and elsewhere. But Handbrake nonetheless pushes all the CPU cores to their restrict, and does in order a check of their extended efficiency.
Note that there are lots of newer and optimized variations of Handbrake which is able to ship improved efficiency, however we use the older model for consistency’s sake. Here, the Desktop nonetheless performs effectively.

Mark Hachman / Foundry
As anticipated, we’ve included the outcomes of two distinction video games to enrich the gaming outcomes we printed above.
By now, Metro: Exodus could also be six years outdated, however the engine punishes outdated and new GPUs alike. It’s value noting that we’ve turned off body era and upscaling for these, offering (within the phrases of my colleague Adam Patrick Murray) “farm to table frames” that measure precisely what these GPUs can render.
The Framework Desktop can’t obtain what most avid gamers take into account to be the minimal 60 fps that makes a recreation playable, and it falls to the underside of the rankings. Still, for an all-around answer like AMD’s Strix Halo chip, it’s spectacular.

Mark Hachman / Foundry

Mark Hachman / Foundry
One of the important thing causes to contemplate a Framework Desktop, nonetheless, is that it lends itself to the usage of AI. Here, the efficiency issues to some extent, but it surely’s the scale of the mannequin that makes it notably enticing. Since the huge quantity of accessible VRAM might be allotted to loading and storing the AI mannequin, our Desktop evaluate unit can load fashions others can’t.
An area AI mannequin is personal, and doesn’t endure from any restrictions on what number of instances you may ask questions. You can obtain no matter fashions will match inside the reminiscence limitation, and practice them to reply your individual private questions.
As it seems, two fashions debuted late in our testing interval: Meta’s Llama 4 Scout 109B (109 billion parameters) mannequin, a 67GB obtain, and the primary OpenAI mannequin to run domestically, referred to as OpenAI-gpt-oss 20B (20 billion parameters), which took up 11.3 GB of area.
When offering a pattern question (“How much wood could a woodchuck chuck if a woodchuck could chuck wood? Be explicit, and let’s pretend that only pine trees are used.”) the Llama 4 Scout mannequin generated 12.64 tokens per second, with 0.62 seconds to first token. OpenAI’s mannequin was a lot sooner, at 53.9 tokens per second, and a pair of.27 seconds to the primary token.
What do these numbers imply? An AI prints textual content like a dot-matrix printer, spitting out line by line. How acceptable the pace is is determined by how shortly you learn: as a quick reader, 12 tokens per second is quick sufficient. The time to first token is how lengthy a mannequin “thinks” earlier than producing its response.
In common, AI fashions with extra parameters have a tendency to provide higher, extra artistic responses, although there’s some variation in that relying on the way it’s skilled. Smaller fashions have a tendency to provide easier responses. I tended to love OpenAI’s responses higher, as Meta’s wanted a little bit of followup at instances.

Mark Hachman / Foundry
Framework Desktop: Conclusion
You will all the time have the ability to put collectively a PC with precisely the elements you want, and aggressively store for each. Laptops are definitely an efficient use of the area, and embrace a show in addition.
But there’s one thing I actually love contained in the Framework Desktop. Yes, Framework soldered down the reminiscence, which is a disadvantage in comparison with, effectively, a lot of the competitors. But I used to be very charmed by the care and element Framework put into the meeting course of, full with the private contact that the tiles provide.
Yet the Desktop isn’t only a gimmick, both. When I reviewed the Asus ROG gaming pill with an identical AMD “Strix Halo” chip inside, I struggled to know why Asus selected a pill over a clamshell, and even higher, a desktop. The Framework Desktop is what I feel all of us needed: a compact, small-form-factor desktop that packs severe energy and can be utilized for productiveness, gaming, and AI alike.
Should I exploit the time period “love” right here? I nearly wish to. Failing that, there’s a way of thoughtfulness that’s rarer in the present day than ever. The Desktop seems like a product designed for somebody that wishes to make that leap into turning into a tech fanatic, with just a bit assist getting over the hump. Could a dad or mum construct a Desktop with their baby? Absolutely.
I very a lot loved my time with the Framework Desktop. If you select to purchase the Framework Desktop, I feel you’ve made a sensible and fulfilling determination.