AMD · Local AI machinePending
Framework Desktop (192 GB)
AMD machine: AMD Ryzen AI Max+ PRO 495, 192 GB memory, 160 GB for the GPU, 273 GB/s memory speed, about 4.2 tok/s on a 70B model (estimate).
82.1Score
- #7 of 41 ranked
- #1 of 18 AMD machines
Why this score
The score adds up these parts. Each part has a weight. A part counts as the average of its sources, pulled toward 80 when there are few sources.
AI power
50% weightcounts as 83.5Average of 2 sources: 85.2
Experts
30% weightcounts as 81.3Average of 4 sources: 81.6
Users
20% weightcounts as 80.0No sources found yet. No data yet. This part counts as 80.
Good: 4 reviews and buyer ratings.
80 is the field average. A part counts as 80 + (average − 80) × N ÷ (N + 1), where N is its number of sources. One source counts half of the gap from 80. No sources count as 80.
Price
€7659 at Framework (frame.work/at) · Maker's EU shop, ships to Austria
'System, Max+ PRO 495 - 192GB' DIY Edition on the Austrian Framework store; pre-order batch, out of stock when read
Prices checked 8 Oct 2026.
Specs
- Chip
- AMD Ryzen AI Max+ PRO 495
- CPU
- 16 cores / 32 threads
- Memory
- 192 GB
- GPU memory
- 160 GBUp to 160 GB can be assigned to the GPU: AMD's figure for the Ryzen AI Max+ PRO 495 (footnote on its Ryzen AI Halo launch blog), also stated by Minisforum, GMKtec and ACEMAGIC. OS not specified; no higher Linux figure documented. Framework publishes no figure for this model. AMD's general 75% Variable Graphics Memory rule would give 144 GB; the chip-specific 160 GB is used because AMD and the box makers state it as the largest share the BIOS allows.
- Memory speed
- 273 GB/s
- FP16 TFLOPS
- 61 TFLOPS
- NPU
- 55 TOPS
- OS
- Linux, Windows
- Software
- ROCm, Vulkan
- Ethernet
- 5 Gbit/s
- Thunderbolt / USB4
- USB4
- QSFP port
- No
- Idle power
- 10.5 W
- Load power
- 164.2 W
- Noise
- 40.1 dB(A)
- Volume
- 4.5 L
- Weight
- 3.1 kg
- SSD
- None included
- Release
- 2025-08
Measured
- Qwen3 30B-A3B, UD-Q4_K_XL
- 66.3 tok/s writing, 604.8 tok/s prompt · llama.cpp Vulkan (llama-bench, fa=1)
- gpt-oss-120b, not labeled in source
- 36.5 tok/s writing, 515.5 tok/s prompt · llama.cpp Vulkan (llama-bench)
- Shisa V2 70B (Llama 3 architecture, stand-in for Llama 3.3 70B), i1-Q4_K_M
- 4.5 tok/s writing, 94.7 tok/s prompt · llama.cpp HIP rocWMMA (llama-bench)
- Llama 4 Scout 109B, not stated
- 12.6 tok/s writing · not stated (PCWorld test)
Notes
Quieter and more efficient than the EVO-X2 and Bosgame M5 in Notebookcheck's tests (23 dB(A) idle, 40 dB(A) load at 15 cm), with a standard mini-ITX layout and 5 GbE. Memory bandwidth limits dense 70B models to about 5 tokens/s, the store prices are high (EUR 3,889 for 128 GB without SSD), and the 192 GB model ships in November.
Power: Phoronix measured 100-104 W average and about 180 W peak at the wall (128 GB unit), Notebookcheck 81.1 W average and 164.2 W maximum (64 GB unit, Prime95/FurMark stress); the table uses Notebookcheck. Weight: maker 3.1 kg, Notebookcheck 3.058 kg. Linux GPU memory: kyuz0 documents up to 124 GiB, Jeff Geerling reports about 108 GB as the stable maximum. gpt-oss-120b speeds vary by source (36 tok/s here, 46 tok/s tg16 from a forum comment); the cited L1T row is labeled 'Framework Strix Halo' without naming the exact model. The 192 GB model uses AMD's figures for the Ryzen AI Max+ PRO 495 (160 GB GPU memory, 273 GB/s, 61 TFLOPS); Framework gives none. GPU memory follows AMD's Variable Graphics Memory rule for every box with this chip (75% of RAM: 48 GB on 64 GB, 96 GB on 128 GB); tuned Linux figures are in the GPU memory note. FP16 TFLOPS is AMD's figure for the chip.
More reading:phoronix.comphoronix.comservethehome.comforum.level1techs.comnotebookcheck.net