Framework Desktop 192GB Costs $6,799; November Batches Sold Out

The Framework Desktop 192GB, built on AMD's Ryzen AI Max+ PRO 495 with 273 GB/s of memory bandwidth, costs $6,799 on pre-order. Framework said on Sep. 30 that shipments start in November, but its store on Nov. 3 showed the first two November batches sold out and a third batch shipping in December. The 128 GB Ryzen AI Max+ 395 model sits out of stock at $3,449.

No outlet has published measurements of the 192 GB PRO 495 yet, so this file covers its specifications and the results for the 128 GB Max+ 395 model. ServeTheHome, which reviewed a Max+ 395 unit, measured roughly 45 tokens per second on gpt-oss-120b MXFP4 in LM Studio with default settings. Phoronix set that model against Dell's GB10 box in llama.cpp and found the Dell faster in every chart on its third page, but published those results only as chart images.

What's new

The 192 GB model changes the chip and the memory, not the chassis. Framework said on Sep. 30 that it pairs the Ryzen AI Max+ PRO 495, with 16 Zen 5 cores and a 40-core Radeon 8065S GPU, with 192 GB of LPDDR5X at 8,533 MT/s on a 256-bit bus, for 273 GB/s. The Max+ 395 model tops out at 128 GB and 256 GB/s, according to Framework's launch post. Both use the same 4.5-litre mini-ITX case. For the 395 model Framework rated 120 W sustained and 140 W boost with a 400 W FlexATX power supply; it states no rating for the 192 GB model.

Price moves with capacity. Framework's store lists the 192 GB system at $6,799 on pre-order, DIY Edition and without storage, and the 128 GB Max+ 395 DIY Edition at $3,449, out of stock, with a 2 TB SSD adding $505. ServeTheHome priced its tested 395 configuration, with a 2 TB SSD and a Noctua fan, at about $2,500 in November 2025. Those figures include different parts, so they do not show how far the price has moved.

What it should solve

Memory capacity sets the largest model that fits, and bandwidth sets how fast each token comes out. Framework says one 128 GB unit runs Llama 3.3 70B at Q6. For the 192 GB model it claims 300B-class models at 4-bit quantization, and its blog names DeepSeek-V4.1-Flash at Q2 and MiMo V2.6 Flash RL at MXFP4 as models the extra memory allows. Those are the maker's claims. The bandwidth gain is smaller than the capacity gain: 192 GB against 128 GB, but 273 GB/s against 256 GB/s. A dense model that fills the larger memory should run slower per token than a smaller one, because each token reads more weights over a bus only slightly faster. No one has measured that yet.

Why Memory Bandwidth Limits Local AI Speed

A language model reads its weights from memory again for every token it generates, so memory bandwidth usually caps speed more than compute does, while capacity decides which models fit at all. Framework says its 192 GB Desktop reaches 273 GB/s, against 256 GB/s for the 128 GB model, so capacity grows far more than bandwidth.

That widens the list of models the machine can load more than it speeds up any one of them. ServeTheHome measured roughly 45 tokens per second on gpt-oss-120b with the 128 GB Max+ 395, and a dense model that fills 192 GB must stream far more data per token over a bus only slightly faster. No outlet has published tokens per second for the 192 GB machine, so that remains an expectation, not a result.

What the 395 measurements show

ServeTheHome's roughly 45 tokens per second on gpt-oss-120b came from an out-of-the-box LM Studio setup, with no context length or batch size stated. It measured 10 to 20 W at idle and 140 to 170 W under load, with the system often sitting at 140 W and running a day and a half at that level without trouble. Noise was 39 to 41 dBA under load with the Noctua fan, which ServeTheHome called quieter than the Minisforum S1-Max. It credited the large heatsink and 120 mm fan.

Its complaints were practical. The kit requires assembly, and ServeTheHome first fitted the Noctua fan in the less effective orientation; the system worked, but sustained loads caused problems and more noise. Wired networking is 5GbE only, which made loading models from network storage slightly slower. There are two front expansion slots, and the board's PCIe Gen4 x4 slot has no rear opening for a card. ServeTheHome ranked it third among the Strix Halo machines it tested, behind the Minisforum S1-Max and Beelink GTR9 Pro, and with a 96 GB GPU allocation the S1-Max ranged from 1% behind to 28% ahead of the Framework in Geekbench AI.

Phoronix's llama.cpp runs, on Jan. 7, used ROCm on the Framework and CUDA on the Dell Pro Max GB10, with gpt-oss-20b, DeepSeek-R1-Distill-Llama-8B and Qwen3-8B at Q8_0. The Dell led every chart on that page and had the better performance per watt. Phoronix wrote that the Framework could at best match it on performance per dollar with GPT-OSS 20B.

How it compares

The closest rival is AMD's own box on the same chip. Voltgrid's AMD Ryzen AI Halo review covers the $3,999 version with a 2 TB SSD, as StorageReview priced it in July, at 256 GB/s. StorageReview's vLLM serving run on gpt-oss-120b reached 222 tokens per second in aggregate at batch 64, against 701 on a DGX Spark. HP's Z2 Mini G1a uses the same chip class; StorageReview measured it close to 40 tokens per second on gpt-oss-120b in LM Studio in a review dated Aug. 8, 2025, with 96 GB assigned to the GPU, a different outlet and setup from ServeTheHome's 45.

The GB10 boxes have the same 273 GB/s as the new 495 model, but only 128 GB. The Dell Pro Precision GB10 review covers the box StorageReview tested as the Dell Pro Max with GB10; Dell.com lists it at $7,967 after moving from $8,103 earlier on Nov. 2. StorageReview measured 68.57 tokens per second on gpt-oss-120b at batch 1 in vLLM. That number used a different tool from ServeTheHome's, so the two do not compare directly. Our GB10 comparison puts GB10 prices between $4,759 and $7,967, with different storage in each box. At $6,799 without storage, the 192 GB Framework sits inside that range, with more memory than the 128 GB GB10 boxes and the same 273 GB/s.

What is not known yet

Three things will decide the 495 model. The first is measured speed on a 120B-class model against the 395's roughly 45 tokens per second. The second is whether models near 192 GB run at usable speed on 273 GB/s. The third is whether the November batches ship on time: Batches 1 and 2 are sold out and Batch 3 is listed for December. Framework has not given a restock date for the 128 GB Max+ 395. Voltgrid will revisit the machine when independent measurements of the 192 GB model are published.