Framework Desktop Review: 128GB Ryzen AI Max+ 395 at 45 tok/s, $3,449

The Framework Desktop with AMD's Ryzen AI Max+ 395 and 128 GB of memory ran gpt-oss-120b at roughly 45 tokens per second for ServeTheHome, but on Nov. 4 Framework's US store listed it at $3,449, system only, and out of stock.

That price is $1,450, or about 73%, above the $1,999 Framework announced for the same 128 GB DIY configuration on Feb. 25, 2025. Every Ryzen AI Max 300 configuration on the store page, from the $1,269 Max 385 with 32 GB to the $1,959 Max+ 395 with 64 GB, showed "out of stock" with a notify-me button. Framework gave no restock date. This review sets the maker's claims against three independent sources: ServeTheHome's review of a 128 GB unit it bought, Notebookcheck's review of a 64 GB unit, and a reproducible llama.cpp sweep that the Level1Techs forum member lhl ran on pre-production 128 GB units and published on GitHub. The 192 GB sibling with the Ryzen AI Max+ PRO 495 is a different machine, covered in our report on its $6,799 pre-order.

What's new

The Framework Desktop is a 4.5-litre mini-ITX desktop, 96.8 x 205.5 x 226.1 mm and 3.1 kg, built around a mobile-class chip. The Ryzen AI Max+ 395 puts 16 Zen 5 cores and a 40-core Radeon 8060S GPU on one package with 128 GB of LPDDR5x-8000 shared between them. Framework says the memory runs on a 256-bit bus for 256 GB/s, and that it had to solder the memory to reach that speed, after months of trying modular options with AMD.

The power budget is a desktop's, not a laptop's. Framework rates the chip at 120 W sustained and 140 W boost, fed by a 400 W FlexATX power supply it developed with FSP. The mainboard is standard mini-ITX with a PCIe x4 slot, two M.2 2280 slots, two USB4 ports, two DisplayPort, HDMI and 5 Gbit Ethernet, so it can move into another case.

The buying model is also unusual for this class. Every system is a DIY Edition: the board, chip, memory and heatsink arrive installed, and the buyer adds the fan, side panel, front tiles, expansion cards, storage and operating system. On Nov. 4 the store listed a 2 TB SanDisk SN7100 SSD at $505 extra and Windows 11 Pro at $199.

What it solves

Memory capacity decides which models load at all. gpt-oss-120b in MXFP4 is a 63.39 GB file in StorageReview's run on HP's Z2 Mini (Ryzen AI Max+ PRO 395), so it fits in 128 GB with room for context. Framework says one unit runs Llama 3.3 70B at Q6 "at real-time conversational speed", and that several units linked over USB4 and 5 Gbit Ethernet can run the full DeepSeek R1 671B. Both are the maker's claims; no source we read measured either.

Bandwidth decides how fast the answer comes out, and here the sweep is the useful evidence. A model reads its active weights from memory for every token it writes, so speed tracks how many bytes each token needs. lhl measured about 215 GB/s of usable GPU bandwidth out of the 256 GB/s theoretical. On that bus, the dense Shisa V2 70B at Q4_K_M, which used about 40.5 GiB of memory at peak, generated 5.0 tokens per second. Qwen 3 30B-A3B, a mixture-of-experts model with 3 billion active parameters, generated 72.0. The 142-billion-parameter dots1, with 14 billion active, reached 20.6.

That split is the practical lesson. The machine holds large models, and it runs them quickly only when few of their parameters are active per token. gpt-oss-120b is such a model, which is why ServeTheHome's 45 tokens per second is possible on a 256 GB/s bus. Llama 3.3 70B at Q6 is dense and larger than the Q4_K_M 70B that lhl measured at 5.0, which suggests it would run slower still, though no one in our sources measured that exact model and quantization.

Pros and Cons

The strongest pro is the combination of speed and calm. ServeTheHome, which bought its 128 GB unit, measured roughly 45 tokens per second on gpt-oss-120b MXFP4 with an out-of-the-box LM Studio setup on Windows 11 Pro, without stating context length or batch size. It measured 10 to 20 W at idle and 140 to 170 W under load, ran the system for a day and a half at 140 W without trouble, and recorded 36 to 37 dBA at idle and 39 to 41 dBA under load with the Noctua fan, quieter than the Minisforum S1-Max. It credited the large heatsink and the 120 mm fan.

Notebookcheck's 64 GB unit told the same story on power and heat. It measured 10.5 W at idle on average, 140.6 W in Cyberpunk 2077 at 1080p Ultra and a 164.2 W peak, against 178.1 W and 214.9 W for the GMKtec EVO-X2 on the same chip. Under Prime95 the CPU boosted to 4 GHz at 115 W, then settled at 3.6 GHz and 100 W, with cores at 50 to 65 °C. Notebookcheck called it quieter, cooler and more efficient than smaller mini PCs with the same processor. Its unit had half the memory of the AI configuration and it ran no language models, so its numbers speak to thermals, not to tokens per second.

The cons are about money and plumbing. ServeTheHome priced its configuration, with a 2 TB SSD, Windows 11 Pro and cosmetic extras, at $2,500 to $2,600 in November 2025, and said that 5GbE should not be the only wired networking on a system at that price; loading models from network storage took longer, so its team reached for other machines. It also found that the board's PCIe Gen4 x4 slot has no rear opening for a card, and that a team member fitted the Noctua fan in the less effective orientation, which caused problems under sustained load. Notebookcheck listed the soldered memory, no 10 Gbps RJ-45 option, Wi-Fi without 320 MHz channels and no PCIe 5 SSD or ECC support among its cons, and noted that hot air exits through the side panel.

lhl's figures carry their own caveat. They come from pre-production units with near-production firmware, Linux kernels 6.15.5 and 6.15.6, ROCm 7.0 nightly builds and llama.cpp build b5863 from July 10, 2025. lhl chose the best backend for each model, mostly Vulkan for generation, and wrote that prompt processing had a lot of performance still on the table. Results on current software may differ.

Who it's for

It fits a developer who wants to run a mixture-of-experts model in the 100B class locally on x86, on Windows or Linux, without a separate GPU: ServeTheHome's 45 tokens per second on gpt-oss-120b is fast enough for interactive use. It fits a buyer who values a quiet desk machine that draws about 140 W under load and can be opened and repaired. It also fits someone who already owns storage, a fan and an operating system licence, since the $3,449 DIY system needs none of Framework's add-ons.

It does not fit a buyer who needs a dense 70B model at interactive speed, given lhl's 5.0 tokens per second at Q4_K_M, or a team that plans to cluster units: 5GbE and USB4 are the only built-in links, and no independent source has measured Framework's multi-unit claim. As of Nov. 4, a buyer could not order one from Framework's US store at all.

How it compares

The closest rivals use the same chip, so price and build decide most of it. Our AMD Ryzen AI Halo review covers AMD's own box with the same Max+ 395, 128 GB and 256 GB/s, which StorageReview priced at $3,999 with a 2 TB SSD in July. Framework's $3,449 system plus its $505 2 TB SSD comes to $3,954 before a fan or expansion cards, within $45 of the Halo's price. The Halo brings single 10GbE where the Framework has 5GbE. StorageReview's Halo figure of 222 tokens per second aggregate at batch 64 came from vLLM serving, a different tool and load from ServeTheHome's single-user LM Studio run, so the two numbers do not compare.

HP's Z2 Mini G1a is the closest like-for-like speed result. Our Z2 Mini review reports StorageReview's close to 40 tokens per second on gpt-oss-120b in LM Studio with 96 GB assigned to the GPU, and HotHardware's about 24 in the same app. That is the same model and the same application as ServeTheHome's 45, but different labs, units and software versions, so the gap is a range across reviews rather than a controlled head-to-head. Notebookcheck measured the Z2 Mini at 195.1 W in the same Cyberpunk 2077 test, above the Framework's 140.6 W.

GMKtec's EVO-X2 listed at $3,499.99 with 128 GB and 1 TB on GMKtec's US store on Oct. 30, with the same 120 W sustained and 140 W peak rating. ServeTheHome ranked the Framework third among the Strix Halo machines it tested, behind the Minisforum S1-Max and Beelink GTR9 Pro and ahead of the EVO-X2, which it called usually the cheapest and likely the best price-performance pick if networking does not matter.

The NVIDIA alternative costs more. Our DGX Spark review covers the 128 GB GB10 box at 273 GB/s, listed at $6,950 with 4 TB of storage in a copy of NVIDIA's marketplace page cached Oct. 31, against $3,449 for the Framework system without storage ($3,954 with a 2 TB SSD), with 200 Gbps ConnectX-7 networking for linking up to four units. ServeTheHome argued the GB10 boxes make sense for buyers who only do AI on NVIDIA's stack, and Strix Halo for those who want Windows, gaming on the side or lower cost. According to our 192 GB report, Phoronix's llama.cpp runs put Dell's GB10 box ahead of the Framework on smaller models. The table below the story sets out every machine's sourced specs.

What to watch

The first date is a restock: Framework's US store showed every Max+ 395 configuration out of stock on Nov. 4 with no date given, and the $3,449 price may move again when it returns. The second is the 192 GB PRO 495 model, whose first two November batches were sold out on Nov. 3, with a third batch listed for December; its first independent measurements will show what 273 GB/s and 192 GB change. The third is software: no recognized lab has published a vLLM serving or multi-user test on a Framework Desktop, and lhl's sweep used mid-2025 nightly builds, so a current llama.cpp or vLLM run on gpt-oss-120b would settle where the machine stands today.