How much does a Strix Halo mini PC cost in 2026?

As of September 2026, the cheapest in-stock 128GB Ryzen AI Max+ 395 box is the GMKtec EVO-X2 at $3,499.99 (128GB/1TB) on GMKtec's own store. The Minisforum MS-S1 MAX is $3,799 (128GB/2TB, in stock), the Beelink GTR9 Pro is $4,349 (128GB/2TB, available), and the Framework Desktop 128GB lists at $3,449 but is out of stock. The $1,999 price from 2025 and early 2026 no longer exists; the DRAM shortage repriced the category.

Is the Ryzen AI Max+ 395 worth it for local LLMs?

Yes, if you run sparse mixture-of-experts models on Linux. Community llama.cpp benchmarks (kyuz0 Strix Halo toolboxes, May 2026) show gpt-oss 120B at about 52-57 tok/s and Qwen3.6-35B-A3B at about 52-60 tok/s. No, if you mainly run dense models: the same benchmarks show a dense Qwen3.6-27B at Q8 decoding at about 6.5 tok/s, because the memory bus moves only 256 GB/s.

What is the best alternative to a Ryzen AI Max+ 395 box?

It depends on what you need. For CUDA, the NVIDIA DGX Spark at $4,699 has 128GB and 273 GB/s. For dense-model speed, the Mac Studio M5 Max with 128GB at $5,099 has 614 GB/s, 2.4 times the Strix Halo bandwidth. For models under 48GB, two used RTX 3090s at about $1,453 each give 936 GB/s per card. Prices are US, as of September 2026.

How much of the 128GB can the Strix Halo GPU use?

AMD states that up to 96GB can be set as dedicated VRAM in the BIOS, and that Windows uses this 96GB Variable Graphics Memory. On Linux, AMD documents a TTM kernel parameter that raises the GPU allocation to 120GB. That makes Linux the better operating system for 100B-class models on this hardware.

Should I wait for the 192GB Ryzen AI Max+ PRO 495 instead?

Only if you need more than 120GB of GPU memory. The PRO 495 adds memory, not speed: vendor figures put bandwidth at 273 GB/s against 256 GB/s. As of 23 September 2026, Minisforum's store shows no price for its 192GB model, and Acemagic lists its F9A PRO 495 at $5,999 without a stated memory configuration and not available to order.

← All guides

Hardware September 23, 2026

Is a Strix Halo Mini PC Worth $3,500 for Local AI?

A Strix Halo mini PC puts 128GB of unified memory on a desk for $3,499.99, the cheapest 128GB box you can buy in September 2026. The '$2,000 Strix Halo' is gone. At $3,500 the question changes: is the cheapest 128GB also the best value? For sparse MoE models on Linux, yes. For dense models, a more expensive Mac is cheaper per unit of speed.

Bottom Line

Price: $3,499.99 for the cheapest in-stock 128GB box (GMKtec EVO-X2, 128GB/1TB), as of September 2026. It is not a $2,000 box any more.
Yes, if you run sparse MoE models on Linux. gpt-oss 120B decodes at about 52-57 tok/s in community benchmarks. No other 128GB box costs less.
No, if you run dense models. A dense 27B at Q8 decodes at about 6.5 tok/s. A $5,099 Mac Studio M5 Max has 2.4 times the bandwidth.
No, if you need CUDA. Buy the DGX Spark at $4,699.
Maybe, on Windows. The GPU gets 96GB, not the 120GB that Linux allows.
Do not wait for the 192GB PRO 495 unless you need more than 120GB of GPU memory. It is not faster.

Already decided? Which Strix Halo box to buy compares the four vendors. Best models for the Ryzen AI Max+ 395 lists what to run on it.

What $3,500 Buys

Every Strix Halo box uses the same chip. AMD lists the Ryzen AI Max+ 395 with 128GB maximum, 256-bit LPDDR5x and LPDDR5x-8000. That bus gives 256 GB/s theoretical bandwidth (our arithmetic: 8,000 MT/s x 32 bytes). The GPU is a 40-core Radeon 8060S.

We checked the vendor stores on 2026-09-23:

Box (128GB)	Price, Sep 2026	Storage	Stock on 2026-09-23
GMKtec EVO-X2	$3,499.99	1TB	In stock (GMKtec store)
GMKtec EVO-X2	$3,649.99	2TB	In stock (GMKtec store)
Framework Desktop	$3,449	none (DIY)	Out of stock
Minisforum MS-S1 MAX	$3,799	2TB	In stock (Minisforum store)
Beelink GTR9 Pro	$4,349	2TB	Available (Beelink store)

The Framework price is lower, but you cannot buy it today. The Framework is also a DIY kit without storage. So the real entry price is $3,499.99.

That price is 75% more than the $1,999.99 GMKtec listed in August. In August that $1,999.99 box was sold out. The soldered LPDDR5X is most of the bill of materials, so these boxes follow the DRAM shortage.

What It Runs

These numbers come from the kyuz0 Strix Halo toolboxes benchmark set. It is community-reported, llama.cpp on Fedora 43, captured 16 May 2026.

Model	Type	Quant	Decode (tg128)	Best backend
gpt-oss 120B	MoE	MXFP4 (59GB)	56.6 tok/s	Vulkan RADV
Qwen3.6-35B-A3B	MoE, 3B active	Q4_K_XL (21GB)	60.4 tok/s	Vulkan RADV
Qwen3.5-122B-A10B	MoE, 10B active	Q5_K_XL (86GB)	22.3 tok/s	Vulkan RADV
Qwen3.6-27B	Dense	Q8_K_XL (33GB)	6.5 tok/s	ROCm 7.2.3
gemma-4-31B	Dense	Q4_K_XL (17GB)	11.5 tok/s	Vulkan RADV

The pattern is simple. Speed depends on active parameters, not total size. The 120B MoE model decodes almost 9 times faster than the 27B dense model at Q8.

A dense 70B at Q4 reads about 40GB per token. At 256 GB/s the ceiling is about 6 tok/s (our arithmetic). Our model guide records about 5 tok/s in practice.

The long-context cost

Short-prompt benchmarks hide one problem. Prompt processing slows as the context fills. The same kyuz0 data shows it for gpt-oss 120B:

Context depth	Prefill, Vulkan	Prefill, ROCm 7.2.3	Decode, Vulkan	Decode, ROCm 7.2.3
Empty	720 tok/s	626 tok/s	56.6 tok/s	51.8 tok/s
32K	308 tok/s	589 tok/s	43.0 tok/s	36.3 tok/s
64K	146 tok/s	371 tok/s	35.0 tok/s	26.1 tok/s

Our arithmetic from those points: filling a 64K context takes about 3.5 minutes on Vulkan and about 2 minutes on ROCm. Prompt caching removes most of that on later turns. A coding agent that sends a new 64K prompt each turn will wait that long each time.

The backend choice also depends on context length. Vulkan wins decode at every depth. ROCm wins prefill at 32K and above, by 1.9 to 2.5 times.

How Much Memory the GPU Gets

AMD’s own developer guide states that the BIOS can dedicate up to 96GB of the 128GB as VRAM. On Linux, a TTM kernel parameter raises the GPU allocation to 120GB. AMD’s Windows driver uses the 96GB Variable Graphics Memory setting.

The difference matters. gpt-oss 120B is 59GB, so it fits either way. Qwen3.5-122B at Q5 is 86GB, which leaves little room for context under the 96GB limit. Run Linux if you bought 128GB for large models.

What the Same Money Buys Instead

Prices are US, as of September 2026. The two ratio columns are our arithmetic from the listed price, total memory, and rated bandwidth.

Option	Price	Memory	Bandwidth	$ per GB	$ per GB/s
Strix Halo (GMKtec EVO-X2)	$3,499.99	128GB	256 GB/s	$27.34	$13.67
Strix Halo (Minisforum MS-S1 MAX)	$3,799	128GB	256 GB/s	$29.68	$14.84
NVIDIA DGX Spark FE	$4,699	128GB	273 GB/s	$36.71	$17.21
Mac Studio M5 Max, 40-core GPU	$5,099	128GB	614 GB/s	$39.84	$8.30
One used RTX 3090 (card only)	~$1,453 avg	24GB	936 GB/s	$60.54	$1.55

The Mac is cheaper per unit of bandwidth than Strix Halo. It costs $1,599 more and gives 2.4 times the bandwidth. On dense models, bandwidth sets decode speed. A dense 70B at Q4 has a ceiling near 15 tok/s on the Mac against about 6 tok/s on Strix Halo (our arithmetic). Mac pricing is from Apple’s configurator on 2026-09-21.

The DGX Spark costs $1,199 more for 7% more bandwidth. You pay for CUDA, not speed. Our DGX Spark verdict explains who needs that.

Two used RTX 3090s cost about $2,906 for the cards. You also need a host PC, which this page does not price. They hold 48GB, so gpt-oss 120B does not fit on the GPUs. For models under 48GB they are much faster. See which used 3090 to buy.

Should You Wait for the 192GB PRO 495?

The Ryzen AI Max+ PRO 495 raises memory to 192GB. It does not raise speed much: vendor figures give 273 GB/s against 256 GB/s. We covered it in PRO 495: wait or buy.

On 2026-09-23 we checked two stores. Minisforum’s MS-S1 MAX-P495 page shows no price and cannot be ordered. Acemagic lists its F9A PRO 495 at $5,999, marked unavailable, with no memory configuration named. Press reports of $7,000-8,000 for loaded units are estimates, not store prices.

Wait only if your target model needs more than 120GB of GPU memory. For gpt-oss 120B and smaller, the 128GB box does the same job now.

The Verdict by Buyer

You are	Worth $3,500?	Buy this instead
MoE user on Linux (gpt-oss 120B, Qwen3.6 MoE)	Yes	Nothing cheaper holds these models
Dense-model user (27B-70B dense)	No	Mac Studio M5 Max 128GB, or used 3090s
CUDA developer	No	DGX Spark
Windows-only user	Maybe	Accept the 96GB GPU limit, or use Linux
Long-context coding agent	Maybe	Use ROCm for prefill and test your context size first
Wants 192GB	Not yet	Wait for a real PRO 495 price

If you are a “yes”, buy the cheapest in-stock 128GB box. The silicon is the same in every one.

GMKtec EVO-X2: the $3,499.99 entry price on GMKtec’s store. Confirm the 128GB SKU before you order.
Minisforum MS-S1 MAX: $3,799 with 2TB and dual 10GbE.
Other Ryzen AI Max+ 395 128GB boxes: check the memory size and stock on each listing.

Amazon prices differ from the vendor-store prices above. Amazon affiliate links: we earn a small commission at no cost to you.