Strix Halo mini PCs for local LLMs

Ranked: EVO-X2, then GTR9 only with a caveat

Q4 fit below uses parameters × 0.5–0.6 GB plus 1–4 GB of runtime and KV cache. MoE models keep every expert in memory. Prices are Amazon observations, not estimates. The EVO-X2 price was $3,649.99 on 1 October and again on the 2 October catalog check. The GTR9 Pro price is the 1 October buy box only.

The default row is the only one I would click without a follow-up question to the seller. The GTR9 link is there because the listing is real and in stock, not because it wins.

GMKtec EVO-X2 — the default

The configuration I tracked is Ryzen AI Max+ 395, 128 GB LPDDR5X, 2 TB SSD, Radeon 8060S, ASIN B0F53MLYQ6, sold by GMKtec-US. Newegg lists the RAM as soldered and a 45–140 W cTDP range. Treat 140 W as the performance-mode ceiling, not a wattage I measured at the wall.

Amazon's product title says "Radeon 8090S". The same listing's body, plus GMKtec and Newegg, describe the Radeon 8060S with 40 CUs. I am cataloging it as the 8060S. If a seller's spec sheet still says 8090S, ask them which silicon is actually in the box. The listing also claims up to 96 GB of the 128 GB pool can be assigned as VRAM. Budget on that split: the OS still needs a slice.

A dense 70B at Q4 is about 35–42 GB of weights before KV cache, so it fits, and at ~256 GB/s it is a slow interactive box — published class figures are around 5 tokens per second, not a 4090. Qwen3.8-Flash-Next is the MoE that belongs here: about 125B total and ~6B active, on the order of 70–80 GB at Q4 before its embedding-table overhead. DeepSeek-V4-Flash-0731 is about 284B total. A normal 4-bit is ~155 GB, which does not fit; an aggressive quant is the honest claim. Peer reports put a gpt-oss-120B-class MoE near 31 tokens per second. I did not re-bench those numbers for this pass. llama.cpp Vulkan, ROCm where the stack supports this APU, and LM Studio are the runners. There is no CUDA. Networking is a single 2.5GbE port, which is fine for pulling a model and serving one household.

128 GB at ~256 GB/s. Confirm the 8060S and that $3,649.99 is still the offer you see.

Beelink GTR9 Pro — same speed, worse default

ASIN B0GQXDCKN1 is a real 128 GB Max+ 395 box. On 1 October 2026 the buy box was $4,349.00, only 3 left, sold by Beelink Direct. That is about $700 over the EVO-X2 for the same CPU, the same 8060S, and the same ~256 GB/s memory. Tokens per second do not improve. What you pay for is dual 10GbE, and that is the part with the failure history.

Version 1 boards ship dual Intel Ethernet Controller E610 10GBASE-T ports. Under sustained network plus iGPU load — an LLM download, a multi-gig model pull, cluster serving — those NICs lock up, drop out of Device Manager or PCI, and on some units bluescreen Windows with SYSTEM_THREAD_EXCEPTION_NOT_HANDLED in ixw.sys. On Linux the E610 path can enter firmware recovery; a full AC power cycle is what brings the port back, not a reboot. Beelink's forum thread "GTR 9 Pro Ethernet Malfunction under load" is the primary owner record. ServeTheHome's review documents the Intel E610 pair, and the comments report crashes under load on Windows and Arch. Craig Wilson's write-up covers the BSOD and firmware lock-up. The Strix Halo wiki marks v1 as a major issue with no reliable software-only fix, and v2.2 as the Realtek swap.

B0GQXDCKN1's page does not say Intel or Realtek. It says "2 x 10Gbps Ethernet." The star line on that ASIN looked family-pooled when I read it — review text cites older GTR units — so I am not treating a high average as evidence this board is fine. Ask the seller, in writing, whether the unit is board v2.2 with Realtek NICs. If they cannot answer, buy the EVO-X2. If you need dual 10GbE and they confirm Realtek, the Q4 model fit is identical to the EVO-X2 and the premium is the networking, not the LLM speed. Stock was three units. That can vanish.

Same 128 GB at ~256 GB/s. $4,349 on 2026-10-01. Not the default.

Older GTR9 listing — watch, do not buy

ASIN B0FPQQYWQ1 is the earlier Crucial 2TB listing. On 1 October 2026 it was currently unavailable, with no featured offer, so I am not publishing a price and I am not putting a button on it. The page copy does say Realtek dual 10GbE. The product-specific rating was 3.4 from 20 reviews, and the review themes are blunt: the 10GbE ports fail, and people return the machine because they cannot pull model weights over those NICs. If it restocks, it is still a caution, and only after the revision is confirmed. The catalog row exists so the ASIN is not forgotten. It is not a recommendation.

SER9 MAX and SER10 MAX — skip for serious LLMs

These are real Beelink boxes, and they are the wrong class. The SER9 MAX is a Ryzen 7 H 255 with Radeon 780M and 64 GB of DDR5-5600 in dual channel, about 89–90 GB/s. The SER10 MAX is a Ryzen AI 9 HX 470 with Radeon 890M and the same memory ceiling. Tom's Hardware measured the SER10 MAX near 89.6 GB/s, against Halo's 256 GB/s. That is roughly a third of the bandwidth. A 10GbE jack, or an OpenClaw sticker with a small Qwen on the SSD, does not change the memory bus.

At that bandwidth, Q4 comfort is about 7–14B. A 27B Q4 is tight and slow. A 70B Q4 is not a serious target. I am not linking them and I am not putting them in the buy catalog. If the price looks like a bargain next to $3,649.99, the bargain is a different computer.

Macs that were actually in stock

On 1 October 2026 the 48 GB Mac mini M4 Pro listings were unavailable, so I am not telling you to buy one. The clean Amazon Macs that day were the Mac Studio M5 Max 36 GB ($2,449, ASIN B0HGKSQMX6), the Mac mini M5 Pro 24 GB ($1,669.99, B0HGGHNQY6), and the Mac mini M4 Pro 24 GB ($1,569.99, B0DLBVHSLD, sold by Amazon.com). Those fit the 7–32B Q4 band. They do not replace 128 GB at ~256 GB/s for a 70B Q4 with a long context, or for a Flash-Next-class MoE. 64 GB and Ultra configure at Apple, with no Amazon ASIN from that check.

Bandwidth on the M4 Pro 24 GB mini is the 273 GB/s class; the M5 Pro listing cites 307 GB/s. The M5 Max Studio has less capacity than the EVO-X2 even when its fabric is quicker. The software path on the Mac is MLX and Metal. The ranked list is in the Apple Silicon guide.

Which one I would actually buy

• Need 70B Q4 or a large MoE in one small Windows/Linux box, and you accept soldered RAM and no CUDA: EVO-X2. Confirm the 8060S and that $3,649.99 is still the offer.
• Need dual 10GbE on that same APU: GTR9 Pro only with a written Realtek / v2.2 answer, and only if the ~$700 gap is worth the ports. Otherwise skip it.
• See an older GTR9 ASIN come back in stock: treat it as the watchlist row, not a new default.
• Looking at a 64 GB SER9 or SER10 because it is cheaper: skip it for serious local LLMs. Hobby 7–14B Q4 only.
• Want macOS and models in the 7–32B Q4 band: the in-stock path on 2026-10-01 was Mac Studio M5 Max 36 GB, then the 24 GB M5 Pro or M4 Pro minis. For 64 GB or Ultra, configure at Apple.

Estimate fit, including context, on the VRAM calculator before you order. A 128 GB sticker does not mean 128 GB is free for weights, and bandwidth — not the sticker — is what sets tokens per second. There is no Beelink SEI12 plus eGPU 4090 in this catalog. That bundle was not on Amazon, and it stays out.

Questions