← All guides

Which Ryzen AI Max+ 395 Mini PC Should You Buy? (August 2026)

Four vendors sell the same AMD Ryzen AI Max+ 395 with the same 128GB of LPDDR5X at the same ~256 GB/s. The price spread is $2,400. Because the silicon is identical, this is not a performance decision at all — it is a decision about stock, I/O, and warranty.

Setting up a Strix Halo box for OpenClaw?

See our AI training options. We'll get 128GB of unified memory running the models you actually need.

Bottom Line (August 2026)

  • The performance answer is “they are the same.” Same APU, same 256-bit LPDDR5X-8000, roughly the same 256 GB/s. No box on this page is meaningfully faster at inference than any other.
  • Best value, when in stockMinisforum MS-S1 MAX (128GB/2TB, launched at $2,299) or the GMKtec EVO-X2 (listed $1,999.99 sale / $2,199.99 regular).
  • Most repairableFramework Desktop, $1,959 for 64GB and $3,449 for 128GB. Everything is documented and replaceable except the memory, which is soldered like everyone else’s.
  • Best I/OBeelink GTR9 Pro, dual 10GbE and dual USB4. At $4,349 it is also the hardest to justify.
  • The real problem is stock, not price. On the day we checked, every Framework configuration was out of stock, every GMKtec configuration on GMKtec’s own store was sold out, and the Beelink was a 35-day pre-sale.
  • Above ~$4,000, stop buying AMD. The ASUS Ascent GX10 is $3,999 for 128GB with CUDA. The AMD price ladder now overlaps the NVIDIA one at the top.

Prices are US, as of August 2026, and move monthly with the DRAM market.

The Decision Nobody Frames Correctly

Most “best Strix Halo mini PC” articles compare benchmark charts. That is the wrong axis. These machines all use one AMD Ryzen AI Max+ 395 (16 Zen 5 cores, Radeon 8060S iGPU) wired to 256-bit LPDDR5X-8000. Bandwidth is the constraint on local inference, and bandwidth is fixed by the platform at roughly 256 GB/s theoretical. Chassis and cooling change sustained clocks at the margin; they do not change the memory bus.

So the real questions are: can you actually buy it, what does it plug into, and who honors the warranty.

The Four Boxes

Box128GB price (Aug 2026)64GB priceNetworkingStock status when checked
Minisforum MS-S1 MAX$2,299 at launch (128GB/2TB)Dual 10GbE, dual USB4 v2 80GbpsIntermittent; Micro Center locations out
GMKtec EVO-X2$1,999.99 sale / $2,199.99 regular$1,999.99 (1TB)2.5GbE, dual USB4All variants sold out on GMKtec’s store
Framework Desktop$3,449$1,9595GbE, dual USB4, PCIe slotOut of stock; refundable pre-order deposit
Beelink GTR9 Pro$4,349 (2TB, was $4,699)Dual 10GbE (Intel E610), dual USB4Pre-sale, ships within 35 days

Two things jump out of that table. The spread between the cheapest and dearest 128GB box is more than $2,000 for identical compute. And the cheaper the advertised price, the less likely you can complete the purchase.

🛒 THE CATEGORY THIS GUIDE IS ABOUT AMD Ryzen AI Max+ 395 · 128GB Unified Memory Box 128GB of LPDDR5X at ~256 GB/s. Check the exact SKU and stock before ordering — this category sells out constantly. Check current price on Amazon →

The stock problem is the story

This category has a habit that wastes people’s money: the widely-quoted price and the buyable price are different numbers.

On GMKtec’s own store, the EVO-X2 shows $1,999.99 on sale against a $2,199.99 regular price — across the 64GB/1TB, 128GB/1TB and 128GB/2TB variants alike — and every one of them is marked sold out or unavailable. Third-party and marketplace listings for the same 128GB/2TB machine have been seen well above $3,600. Framework lists $1,959 for 64GB and $3,449 for 128GB and sells both as pre-orders with a refundable deposit, not as stock. Beelink’s GTR9 Pro is a pre-sale with a 35-day ship window.

The practical rule: treat any Strix Halo price you see as an offer to join a queue. Before you commit, confirm the exact SKU, the storage tier, and whether the seller has units in hand. A $1,999 machine that ships in five weeks is not competing with a $2,299 machine you can have on Tuesday.

Why these boxes cost double what guides say

The soldered 128GB of LPDDR5X is most of the bill of materials, so these machines track the memory market rather than the silicon market. The Framework Desktop 128GB launched at $1,999 and lists at $3,449 today — up about 73% on memory and storage cost alone, with no change to the chip.

That is why you should distrust any 2026 buying guide quoting $1,199 to $1,899 for a 128GB Strix Halo box. Those are 2025 numbers. The DRAM shortage repriced this entire category, and nothing in the supply forecast suggests relief before 2027.

What they actually run

Capacity is not speed. 128GB of unified memory holds a dense 70B at Q4 with room to spare, and then decodes it at roughly 5 tokens per second — the ~256 GB/s bus is the ceiling, and no chassis fixes that. We measured the same wall when we ranked the cheapest ways to run a 70B.

Sparse MoE models are what this hardware is for. Qwen3-30B-A3B class models run near 100 tok/s and gpt-oss 120B lands in the 31-55 tok/s range on these boxes — see our model picks for the 395 for the full table. Buy one of these because you want a 120B-class MoE model on your desk at low power. Do not buy one to run dense Llama 70B.

Which One Should You Buy?

  1. You want the most memory per dollar and can wait for stockMinisforum MS-S1 MAX or GMKtec EVO-X2, whichever is genuinely in stock the day you buy. Verify the storage tier; the price gap between 1TB and 2TB SKUs is large.
  2. You want to be able to fix it in five yearsFramework Desktop. It is the only box here with a PCIe slot and public repair documentation. You pay roughly $1,000 over the value picks for that, and the RAM is still soldered.
  3. The box is going in a rack or a homelab with 10GbEBeelink GTR9 Pro, but only if the dual 10GbE genuinely earns its keep. Otherwise see the next point.
  4. Your budget is over $4,000 — do not buy an AMD box at all. The ASUS Ascent GX10 is $3,999 for the 1TB model with 128GB unified memory, GB10 Grace Blackwell silicon, CUDA and DGX OS. Buying a $4,349 Strix Halo box in August 2026 means paying NVIDIA prices for ROCm software support.
  5. You need dense-model speed, not capacity — none of these. Go to dual RTX 3090s or a Mac with real bandwidth.

128GBMinisforum MS-S1 MAX 128GB/2TB ↗

128GBGMKtec EVO-X2 ↗

128GBBeelink GTR9 Pro 128GB/2TB ↗

128GBASUS Ascent GX10, GB10 128GB unified ↗

One honest caveat on all of it: the soldered memory that makes these boxes expensive is also what makes them a hedge. You cannot upgrade later at any price, but you also are not exposed to buying DDR5 at four times its old price in a year. Buy the 128GB configuration the first time.

See Also

Need OpenClaw fixed live?

Remote rescue sessions for gateway, auth, tunnel, VPS, and model access problems.

See Rescue Session

Read next

DGX Spark vs Strix Halo for Local LLMs: $4,699 vs $2,600 128GB Boxes
Compare NVIDIA DGX Spark and AMD Strix Halo (Ryzen AI Max+ 395) 128GB mini PCs for local LLMs: bandwidth, CUDA vs ROCm, MoE performance, and pricing.
Strix Halo vs Mac Studio M4 Max 128GB for Local LLMs: Which Unified Memory Box?
Compare AMD Strix Halo (Ryzen AI Max+ 395) and Mac Studio M4 Max 128GB for local LLMs: 256 vs 546 GB/s bandwidth, decode speed, ROCm vs MLX, August 2026 prices.
Jetson Thor vs DGX Spark (August 2026): Which NVIDIA 128GB Box Is For You?
Jetson AGX Thor and DGX Spark both carry 128GB of LPDDR5X at exactly 273 GB/s, so they generate tokens at the same ceiling. Thor lists at $3,499 against Spark's $4,699. The real decision is deploy versus develop, not TOPS — and street pricing reverses the MSRP gap.
MacBook Pro M4 Max for AI: 36GB vs 128GB (Which RAM for Local LLMs?)
36GB or 128GB M4 Max for local AI? The 36GB config ships on the 14-core M4 Max at 410 GB/s; 128GB requires the 16-core chip at 546 GB/s. 36GB runs Qwen 3.6 27B Q8 and Laguna XS 2.1; 128GB is the only way to run gpt-oss 120B or Llama 4 Scout locally.