Ryzen AI Max+ 395 128GB Price & Used Prices: Which Local LLMs Can It Run?

The Ryzen AI Max+ 395 128GB is AMD's 2025 mini PC with shared memory, running a 176B MoE at 4-bit at about 41 tokens/s. You set the GPU's share.

eBay

Ryzen AI Max+ 395 128GB price history

$4,464+1.80%

eBay

Price updated 10/21/2026, 02:20

$7,460$6,305$5,150$3,995$2,840

Low $3,500High $6,800

Sep 12Oct 1Oct 21

As an eBay Partner, VRAMGlass may be compensated if you buy through the eBay links on this page. The price you pay stays the same.

Full Ryzen AI Max+ 395 128GB specifications

Memory128 GB · LPDDR5X · unified memory
Bandwidth256 GB/s
FP16 / BF1659.4 TFLOPS
INT859.4 TOPS

Bus width: 256-bit
Architecture: Strix Halo (Radeon 8060S, RDNA 3.5)
SoC: AMD Ryzen AI Max+ 395
Cores: 2,560
Tensor cores: 80
Ecosystem: ROCm
Power: 120 W
Type: Unified memory
Launch date: Mar 1, 2025

Best Ryzen AI Max+ 395 128GB deals on eBay

Hand-picked deals from trusted eBay sellers. Shop with confidence.

Updated 2 hours ago2 hr. ago

As an eBay Partner, VRAMGlass may be compensated if you buy through the eBay links on this page. The price you pay stays the same.

See more on eBay

What LLMs can a single Ryzen AI Max+ 395 128GB run?

A single AI Max+ 395 is best with Qwen3.8 Flash Next at 4-bit. Windows caps the GPU at 96GB, so 2-bit is safer; Linux can fit 300B-class at 2-bit.

Model	Size	Runs?	Speed t/s	Max context
Gemma 4 31B30.7B	11.8 GB	Runs	15.3	256K
Qwen3.6 35B A3B36B (3B active)	12.3 GB	Runs	96.6	256K
Qwen3.8 Flash Next176B (6B active)	78.9 GB	Runs	52.9	256K
DeepSeek V4 Flash284B (13B active)	96.8 GB	Runs	36.5	1M
Hy3295B (21B active)	107.2 GB	Runs	19.9	86K
GLM 5.3 Flash320B (18B active)	108.7 GB	Runs	27.7	1M

Model	Size	Runs?	Speed t/s	Max context
Gemma 4 31B30.7B	18.3 GB	Runs	10.2	256K
Qwen3.6 35B A3B36B (3B active)	22.1 GB	Runs	68.9	256K
Qwen3.8 Flash Next176B (6B active)	111.3 GB	Runs	40.8	256K
DeepSeek V4 Flash284B (13B active)	155.1 GB	Too big	—	—
Hy3295B (21B active)	182.2 GB	Too big	—	—
GLM 5.3 Flash320B (18B active)	199.7 GB	Too big	—	—

Model	Size	Runs?	Speed t/s	Max context
Gemma 4 31B30.7B	32.6 GB	Runs	5.8	256K
Qwen3.6 35B A3B36B (3B active)	36.9 GB	Runs	48.1	256K
Qwen3.8 Flash Next176B (6B active)	188.2 GB	Too big	—	—
DeepSeek V4 Flash284B (13B active)	161.9 GB	Too big	—	—
Hy3295B (21B active)	317.7 GB	Too big	—	—
GLM 5.3 Flash320B (18B active)	341 GB	Too big	—	—

RunsFits fully in VRAM; speed is estimated from VRAM bandwidth.

Too bigWon’t fit on one card at this quantization, even with RAM offload.

Full calculation rules are on the Methodology page.

Who it suits

Good fit

Running 100B-class MoE models on one machine
Qwen3.8 Flash Next runs at about 41 tokens/s at 4-bit and about 53 at 2-bit
Running DeepSeek V4 Flash at 2-bit
At 96.8 GB, it fits on Linux once you raise the shared memory limit, running at about 37 tokens/s with up to 1M of context
You want a small, low-power x86 machine
The chip draws between 45W and 120W, with official ROCm support on both Linux and Windows

Not a good fit

Mainly running dense models
Qwen3.8 27B at 4-bit runs at about 12 tokens/s, only 30% of an RTX 3090's speed
Running DeepSeek V4 Flash or GLM 5.3 Flash at 4-bit
They need 155 to 200 GB, which 128GB can't hold
Adding memory later
The RAM is soldered to the motherboard and can't be upgraded

Before you buy

Windows: set aside dedicated VRAM in Adrenalin
On Windows, set dedicated VRAM on the Performance → Tuning page of AMD's Adrenalin software, then reboot. The 128GB model can set aside up to 96GB, and AMD says the GPU can then use about 112GB in total (dedicated plus shared). So Qwen3.8 Flash Next at 2-bit fits in 96GB without trouble, while the 4-bit version exceeds 96GB and has to lean on shared memory, with very little headroom.
Linux: set VRAM small in the BIOS, raise the shared limit
On Linux, the GPU mainly uses memory shared with the system, and the default limit is about half of total memory. AMD recommends setting dedicated VRAM in the BIOS very small (0.5 GB, for example), then raising the shared limit with the amd-ttm tool; the change takes effect after a reboot. There are kernel requirements too: Ubuntu 24.04 needs HWE 6.17.0-19 or OEM 6.14.0-1018 or later, and other distributions need 6.18.4 or later, or compute jobs may fail to start.
Software status
ROCm 10.0 supports this chip on both Ubuntu and Windows 11, and Ollama's Linux version recognizes it directly. The vLLM documentation lists it too, requiring ROCm 7.0.2 or later. For llama.cpp, both the ROCm and Vulkan backends work.
Same chip, but systems differ in cooling and power limits
AMD lets manufacturers configure this chip anywhere between 45W and 120W, and machines such as the Framework Desktop, GMKtec EVO-X2 and HP Z2 Mini G1a differ in their settings and cooling. Generation speed depends mainly on memory bandwidth, so it is similar across machines; prompt processing is compute-heavy, so machines with lower power limits will be slower. Before buying, check the sustained power setting of the specific machine.

Compared with similar machines

Ryzen AI Max+ 395 128GB vs DGX Spark 128GB and Mac Studio M4 Max 128GB: price, memory, bandwidth, compute and local LLM speed compared.

Spec	Ryzen AI Max+ 395 128GB	DGX Spark 128GB	Mac Studio M4 Max 128GB
Memory	128 GB	128 GBSame	128 GBSame
Bandwidth	256 GB/s	273 GB/s7% higher	546 GB/s113% higher
BF16 / FP16 compute	59 TFLOPS	125 TFLOPS110% higher	—
INT8 compute	59 TOPS	250 TOPS321% higher	—
Qwen3.8 27B speed (2-bit)	19 t/s	20.2 t/s6% faster	39.3 t/s107% faster
Largest dense model (4-bit)	222B	222BSame	222BSame
Power	120 W	240 W100% higher	480 W300% higher
Released	2025-03	2025-10	2025-03
eBay median	$4,464Buy Ryzen AI Max+ 395 128GB	$8,081Buy DGX Spark 128GB	$5,800Buy Mac Studio M4 Max 128GB