By LLM Hardware team · Methodology ·
Checked 2026-10-13. Author: LLM Hardware team. This site does not store or show Amazon prices. Check current price on Amazon.
Key fact: Qwen3.8-Flash-Next UD-Q4_K_XL is 111.33 GB. A 128 GB Strix Halo box can hold UD-IQ4_XS at 93.68 GB. UD-Q4_K_XL is tight. A 32 GB GPU does not fit either file.
How we got it: Unsloth Qwen3.8-Flash-Next-GGUF shard sums on 9 October 2026. VRAM formula on 180B total (125B plus 51B n-gram embedding and 4B MTP) is 180 × 0.6 + 2 = 110 GB. Bandwidth is the ~256 GB/s LPDDR5X class.
Checked:
Methodology — parameter math, quantization bytes, and the source list.
The brief here is one box that can hold a 70B Q4 model without a second GPU. AMD's Strix Halo (Ryzen AI Max+ 395) is the x86 answer: a 16-core Zen 5 APU, a 40-CU Radeon 8060S, and 128 GB of soldered LPDDR5X-8000 on a 256-bit bus, about 256 GB/s theoretical. A third-party rocm_bandwidth_test run peaked at about 212 GB/s. That is the whole purchase. RAM capacity without that bandwidth is a different machine. A 64 GB dual-channel DDR5 mini PC at ~90 GB/s is not a cheaper Halo.
Q4 fit uses params × 0.6 + 2, the same estimate as the VRAM calculator. MoE models keep every expert in memory. Size LFM2.5-24B-A2B on the LFM hardware page. On 9 October that ASIN had loaded a different PC.
| Rank | Machine | Unified memory | Bandwidth | Q4 model-size fit | Price / stock |
|---|---|---|---|---|---|
| 1 · Buy | GMKtec EVO-X2 EVO-X2 on Amazon | 128 GB LPDDR5X-8000 soldered (up to 96 GB GPU alloc) | ~256 GB/s theoretical (about 212 GB/s peak in a third-party ROCm test) | Flash-Next UD-IQ4_XS is 93.68 GB and fits. UD-Q4_K_XL is 111.33 GB and is tight. Dense 70B Q4 is 44 GB and is slow at ~256 GB/s. | The captured title did not include 8090S. This pass did not re-read a GPU name from the title. Check current price on Amazon. |
| 2 · Caution | Beelink GTR9 Pro GTR9 Pro 128GB on Amazon | 128 GB LPDDR5X-8000 soldered | ~256 GB/s theoretical (same silicon as the EVO-X2) | Same memory class as the EVO-X2: UD-IQ4_XS at 93.68 GB fits, UD-Q4_K_XL at 111.33 GB is tight. Buy it only if you accept the NIC caveat and an out-of-stock listing. | Check current price on Amazon. |
| 3 · Watchlist | Beelink GTR9 Pro (older listing) | 128 GB LPDDR5X | ~256 GB/s theoretical | Same compute tier when it is actually stocked. Not a buy today. | No buy box on 9 October 2026. Title still matched. Unavailable. 3.4★ / 20. |
| Skip | Beelink SER9 MAX / SER10 MAX 64GB | 64 GB DDR5-5600 dual-channel | ~89–90 GB/s (SER10 MAX ~89.6 GB/s) | Hobby ~7–14B Q4 only. 27B Q4 is tight. Not a Halo peer. | Not cataloged. A lower price does not buy bandwidth. |
The EVO-X2 is the in-stock buy. The GTR9 link is the same memory class and was temporarily out of stock on 13 October 2026. The NIC caveat still applies. Links to Amazon are affiliate links.
The configuration this guide tracks 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 listed performance-mode ceiling, not a wall-power reading.
On 13 October 2026 the captured Amazon title did not include "Radeon 8090S". This pass did not re-read a GPU name from that title. Earlier notes, plus GMKtec and Newegg, describe the Radeon 8060S with 40 CUs. 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. LFM2.5-24B-A2B at BF16 is 59.6 GB by our VRAM formula and fits this 128 GB box.
A dense 70B at Q4 is 44 GB by our VRAM formula, so it fits this 128 GB pool, and at ~256 GB/s it is a slow interactive box — published class figures are around 5 tokens per second, not a 4090. Do not read the calculator’s 128 GB ceiling (about 210B) as an interactive dense model at this bandwidth. Qwen3.8-Flash-Next is the MoE that belongs on this box, and it is larger than the old 70–80 GB note. The card is 125B with about 6B active, plus a 51B n-gram embedding and a 4B MTP head. Counting those as 180B, the formula is 110 GB. Unsloth’s UD-IQ4_XS file summed to 93.68 GB on 9 October 2026 and fits. UD-Q4_K_XL summed to 111.33 GB and is tight once the OS and KV cache need room. The 70–80 GB range matches the IQ1 and Q2 files, not Q4. File sizes and the loader command are in the Flash-Next hardware guide. DeepSeek-V4-Flash-0731 is about 284B total. A normal 4-bit is ~155 GB, which does not fit. Peer reports put a gpt-oss-120B-class MoE near 31 tokens per second. Those are peer-reported figures, not independently checked here. 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.
See the GMKtec EVO-X2 on Amazon128 GB at ~256 GB/s. The captured title did not include 8090S. Check current price on Amazon.
On 13 October 2026 ASIN B0GQXDCKN1 loaded as the Beelink GTR9 Pro 128G LPDDR5X, sold by Beelink Direct, temporarily out of stock. The order note said you are not charged until it ships. The 11 October title matches the 128 GB machine, so the link is back, and the EVO-X2 remains the in-stock default. Tokens per second match the EVO-X2 when the box is a Max+ 395. Dual 10GbE does not add memory bandwidth, and that NIC is the part with the failure history. Check current price on Amazon.
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.
Earlier reads of a GTR9 listing said "2 x 10Gbps Ethernet" and did not print Intel or Realtek. A star average is not evidence the board is fine. If a seller offers you a GTR9, ask in writing whether the unit is board v2.2 with Realtek NICs. If they cannot answer, buy the EVO-X2. The button below is the 128 GB listing that matched on 13 October, and it was temporarily out of stock.
13 October 2026: title matched the GTR9 Pro 128GB, Beelink Direct, temporarily out of stock. 9 October: the same ASIN loaded a SER3. Links to Amazon are affiliate links.
GTR9 Pro 128GB on Amazon — out of stockASIN B0FPQQYWQ1 is the earlier Crucial 2TB listing. On 13 October 2026 the title matched the GTR9 Pro 128 GB Max+ 395. Add to cart was present and the availability line was blank, so this page does not claim it is in stock and does not add a buy button. The NIC caution still applies. This listing is not the recommendation. Prefer the in-stock EVO-X2. Check current price on Amazon.
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's SER10 MAX review gives its memory bandwidth cap as 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. This guide does not link them or put them in the buy catalog. Check current price on Amazon.
Those Macs fit the 7–32B Q4 band. LFM2.5-24B-A2B Q4 is 16.4 GB and is tight on 24 GB unified. They do not fit Qwen3.8-Flash-Next. The M5 Ultra 96 GB base fits a 70B Q4 file (44 GB) and does not fit Flash-Next (110 GB). It had no verified US dollar and no Amazon ASIN. The ranked Mac list is in the Apple Silicon guide.
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.
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.
The GMKtec EVO-X2 128GB (ASIN B0F53MLYQ6) is the default because it was in stock. The captured title did not include 8090S. It is a Ryzen AI Max+ 395 with about 256 GB/s and a single 2.5GbE port. Qwen3.8-Flash-Next UD-IQ4_XS is 93.68 GB and fits. UD-Q4_K_XL is 111.33 GB and is tight. LFM2.5-24B-A2B BF16 formula is 59.6 GB and fits. The availability line was blank, so stock is not claimed. Do not copy that dollar onto B0GQXDCKN1. Check current price on Amazon.
Q4 on this site is params × 0.6 + 2. Mixture-of-experts models occupy their total parameter count, plus extra tables when the card lists them. A dense 70B Q4 is 44 GB and fits 128 GB, and it is slow at about 256 GB/s. Qwen3.8-Flash-Next counted as 180B (125B plus the 51B n-gram embedding and 4B MTP) is 110 GB. Unsloth UD-IQ4_XS is 93.68 GB and fits. UD-Q4_K_XL is 111.33 GB and is tight. A 24 GB or 32 GB GPU does not fit those files. There is no CUDA.
No. Both use a Ryzen AI Max+ 395, a Radeon 8060S, 128 GB of soldered LPDDR5X-8000, and about 256 GB/s theoretical bandwidth. Decode speed is the same class. The GTR9 Pro costs more and adds dual 10GbE, which does not raise memory bandwidth.
Version 1 boards use dual Intel E610 10GbE controllers that lock up under combined GPU and network load — the exact pattern of an LLM download or a cluster pull. Symptoms include the NIC disappearing, a Windows BSOD in ixw.sys, and a Linux firmware-recovery state that needs a full AC power cycle. Board v2.2 replaces Intel with Realtek. On 13 October 2026 ASIN B0GQXDCKN1 titled as the GTR9 Pro 128GB, Beelink Direct, temporarily out of stock. On 9 October that ASIN had loaded a SER3. ASIN B0FPQQYWQ1 still titled as the GTR9 Pro. Prefer the in-stock EVO-X2, and still ask whether a GTR9 is board v2.2. Check current price on Amazon.
No. Skip them for serious local LLMs. They are 64 GB dual-channel DDR5 boxes around 89–90 GB/s, about one-third the bandwidth of Strix Halo. That is a hobby tier for roughly 7–14B at Q4. A 27B Q4 is tight and slow. A 10GbE port does not change that.
Neither fits Qwen3.8-Flash-Next. LFM2.5-24B-A2B Q4 is 16.4 GB: tight on the 24 GB mini after a macOS reserve, and it fits the 36 GB Studio. The M5 Ultra 96 GB base configures at Apple, with no verified US dollar. A 70B Q4 file (44 GB) fits that 96 GB base. Flash-Next at 110 GB does not. Those Macs are a different bandwidth tier from 128 GB Strix Halo. Check current price on Amazon.