Zotac's Strix Halo and RTX 5060 Ti Mini-PCs in 2026: Two Compact Local AI Paths, Not One Hybrid Box

By RunAIHome Team · 12 min read
mini-pcstrix-halortx-5060-tilocal-llmhardware

TL;DR: The viral headline reads like Zotac built one mini-PC with a Strix Halo APU and an RTX 5060 Ti. It didn’t — those are two separate machines. The Magnus EAMAX395C is a 128GB unified-memory Strix Halo box with no price yet; the Magnus EN275060TC is a 2.65-liter Intel box with a desktop RTX 5060 Ti 16GB, on sale now at $1,599.99 barebone.

Magnus EAMAX395CMagnus EN275060TCGMKtec EVO-X2 128GB
Best for70B–120B models in one small boxFastest responses on ≤14B modelsSame 70B–120B job, buyable today
Price / CostUnpriced in the US (Oct 2026)$1,599.99 barebone + RAM/SSD$3,499–$3,649 complete
The catchYou can’t buy it yet16GB VRAM wall; dense 27B won’t fit256 GB/s bandwidth, same as EAMAX

Honest take: If you want the big-model unified-memory path, buy a GMKtec EVO-X2 today instead of waiting for an unpriced Zotac with the identical chip. If your models stop at 14B, the Magnus EN is the smallest genuinely fast box you can buy — just budget another ~$970 for RAM and storage before it boots.

Zotac’s Computex preview finally turned into shipping products this fall, and the coverage compressed three different machines into one confusing headline. NotebookCheck’s “first AMD Strix Halo mini-PC with RTX 5060 Ti and RTX 5070 Ti options in tow” describes the lineup, not a single computer. Since a Strix-Halo-plus-discrete-Blackwell hybrid would genuinely change the mini-PC math for local AI — 128GB of slow unified memory and 16GB of fast GDDR7 in one chassis — it’s worth being precise about what you can actually order. Before buying anything here, run your target model through our VRAM calculator to see which memory pool it needs.

What Zotac actually announced: three machines, zero hybrids

The Magnus family now splits three ways:

ModelCPU/APUGPUMemoryBandwidth (GPU pool)VolumePSUUS price (Oct 2026)
Magnus EAMAX395CRyzen AI Max+ 395 (16C Zen 5)Integrated Radeon 8060S (40 CU)128GB LPDDR5X-8000, soldered256 GB/s2.65 Lexternalnot published
Magnus EN275060TCIntel Core Ultra 7 255HX (20C)Desktop RTX 5060 Ti 16GB GDDR72× DDR5 SODIMM, up to 96GB448 GB/s2.65 L330W$1,467.99–$1,599.99 barebone
Magnus One EU27507TCIntel Core Ultra 7 265Desktop RTX 5070 Ti 16GB GDDR72× DDR5 DIMM896 GB/s8.48 L650W Platinum~$2,970–$3,000 barebone

The EAMAX395C is pure Strix Halo: the Radeon 8060S integrated GPU does all the inference work against that 128GB LPDDR5X pool (Windows caps the GPU-addressable share at 75%, so 96GB in practice — AMD’s own Variable Graphics Memory FAQ confirms the ceiling). There are also EAMAX390C and EAMAX385C variants with the cut-down Radeon 8050S and 32GB soldered — skip both for AI work; 32GB of 256 GB/s memory is the worst of both worlds. The interesting config is the 395 with 128GB, three M.2 slots, dual 2.5GbE, USB4, and Wi-Fi 7 in the same 2.65-liter chassis as the EN.

The EN275060TC is the one that made “world’s smallest PC with a desktop RTX 5060 Ti” headlines, and the desktop qualifier matters: this is the full 448 GB/s, 16GB GDDR7 card, not the gimped laptop variant. It shipped to the Zotac Store at $1,599.99 barebone (CDW has it at $1,467.99), with a pre-configured Windows version that bundles a useless-for-AI 16GB of RAM.

The Magnus One is the same idea scaled up — desktop RTX 5070 Ti 16GB at 896 GB/s in 8.5 liters for roughly $3,000 barebone. Twice the bandwidth of the EN, same 16GB ceiling, nearly twice the price. For local AI specifically, that’s a hard sell: you’re paying ~$1,400 more for speed on models that were already fast, while the VRAM wall that actually constrains you doesn’t move an inch.

The real decision: a 96GB slow pool or a 16GB fast pool

Strip away the chassis marketing and this is the same fork we’ve mapped in Strix Halo 64GB vs 128GB and the GPU buying guide: decode speed is memory bandwidth divided by bytes read per token, so the two machines aren’t competing — they’re answering different questions.

What the EN’s 16GB fast pool runs well. On a desktop RTX 5060 Ti 16GB, Markaicode’s Ollama benchmarks and the aggregated figures we’ve verified across this site put real decode speeds at:

  • 7B–8B dense, Q4_K_M: 41–51 tok/s
  • 14B dense, Q4_K_M: ~31–35 tok/s
  • Gemma 4 26B-A4B QAT (MoE, ~15GB): ~40–50 tok/s

That last line is the card’s party trick. The QAT build of Google’s 26B MoE squeezes under the 16GB ceiling and decodes like a small model because only ~4B parameters are active per token. One owner has even pushed 256K context on the smaller IQ3_XXS file at 99 tok/s — a stunt, not a daily driver, but it shows the headroom. Full details in best local LLM for 16GB VRAM.

What it can’t do. A dense 27B at Q4_K_M is a ~17GB file. It does not fit in 16GB, and no setting fixes that — Ollama will silently split it across the GPU and CPU, and your 40 tok/s card becomes a 6 tok/s card. This is the most common disappointment we hear from 16GB buyers, and it looks like this:

$ ollama run qwen3.6:27b "explain this error"
# ...glacial output...
$ ollama ps
NAME          SIZE     PROCESSOR          UNTIL
qwen3.6:27b   19 GB    22%/78% CPU/GPU    4 minutes from now

Anything other than 100% GPU in that PROCESSOR column means you’ve hit the wall. The fix on this hardware is to switch to a model built for the tier — ollama run gemma4:26b-a4b-it-qat gets you 26B-class knowledge at ~40–50 tok/s fully resident — or quantize the KV cache (OLLAMA_KV_CACHE_TYPE=q8_0) if you’re only a gigabyte over at long context.

What the EAMAX’s 96GB slow pool runs. Identical silicon to every other Ryzen AI Max+ 395 machine we’ve tested, so the numbers carry over directly from our Strix Halo deep dive:

  • Qwen3-30B-class MoE: ~100 tok/s
  • Qwen3-Next-80B-A3B, Q4: ~59–63 tok/s
  • gpt-oss-120b (MoE, ~61GB resident): 31–56 tok/s depending on stack (31 via stock Ollama on Windows, mid-50s via tuned llama.cpp Vulkan on Linux)
  • Llama 3.3 70B dense, Q4: ~4–6 tok/s — dense models read the whole 42.5GB per token, and 256 GB/s divided by 42.5GB is a hard ceiling of 6

Same lesson as always with this platform: it’s a capacity play, not a speed play. MoE models with small active parameter counts are spectacular; big dense models remain a slideshow.

The price problem, in both directions

The EN barebone trap. $1,599.99 buys you a machine that will not boot: no RAM, no SSD, no OS. In the middle of the worst DRAM pricing crisis in a decade, that’s not a rounding error. A 64GB DDR5 SODIMM kit (2×32GB) runs $550–$1,060 at US retail right now — the mainstream Crucial DDR5-5600 kit sits at $630 at Best Buy — and a 2TB Samsung 990 Pro is ~$338. Realistic all-in cost for a sensible 64GB/2TB build: ~$2,550. For comparison, the bare RTX 5060 Ti 16GB card streets at $679–$805 (MSRP $429 — nobody pays MSRP this year), so Zotac’s packaging premium for the 2.65-liter form factor is real but not absurd. You’re paying for the shoebox, the 330W integrated PSU, Thunderbolt 4, and Wi-Fi 7.

One genuine caution: 16GB VRAM plus 330W total system power means this box runs the 5060 Ti at the edge of its thermal envelope under sustained inference. Reviews so far are gaming-weighted; if you plan to batch-generate overnight, wait for sustained-load data or point a fan at it.

The EAMAX pricing vacuum. Zotac has published full specs, product pages, and a “bring your own storage” barebone SKU — and no US price, nine months after the January announcement. The only listing we found anywhere is a Finnish aggregator showing €6,628, which reads as a placeholder, not a street price. The comparables tell you what to expect: the GMKtec EVO-X2 128GB sells for $3,499–$3,649, the Framework Desktop 128GB is $3,449, and the Beelink GTR9 Pro has drifted to $4,349. Every 128GB Strix Halo machine that has entered the US market in 2026 has landed between $3,400 and $4,400 — and with LPDDR5X caught in the same DRAM squeeze as everything else, Zotac launching below that band would be a surprise.

Which makes the waiting math simple: the EAMAX will be the same chip, same 256 GB/s, same tok/s as machines you can order tonight. Unless Zotac undercuts the market by hundreds of dollars — nothing in its history suggests it will — there is no payoff for waiting on an unpriced SKU.

What to actually buy

Prices as of October 2026, all verified above:

Your situationThe machinePriceWhere
You want 70B–120B MoE models in a small quiet box, todayGMKtec EVO-X2 128GB$3,499–$3,649Check price
Your models stop at 14B (or Gemma 4 26B-A4B) and desk space is the constraintZotac Magnus EN275060TC + 64GB/2TB~$2,550 all-inCheck price
You want the most tok/s per dollar and don’t care about sizeUsed RTX 3090 24GB in any tower$1,150–$1,350 (card)Check price
Undecided — test your actual workload firstRented RTX 3090from $0.07/hrVast.ai

The used RTX 3090 row keeps winning these tables for a reason: 936 GB/s and 24GB means it runs everything the EN runs, faster, plus the dense 27B class the EN can’t hold — for half the all-in money. The full three-year math is in RTX 5060 Ti 16GB vs used RTX 3090. The mini-PCs earn their premium only when a tower is genuinely off the table.

If you end up with any of these boxes as a coding backend, our sister site covers wiring local models into Cline, and aifoss.dev’s Ollama review covers the serving stack all three machines would run.

FAQ

Is there any mini-PC that combines Strix Halo unified memory with a discrete NVIDIA GPU? Not from Zotac, and we’re not aware of any shipping machine that does it. The closest real-world equivalent is a Strix Halo box with an eGPU over USB4/Oculink — possible, fiddly, and bandwidth-constrained on the interconnect.

Does the Magnus EN’s RTX 5060 Ti perform like a desktop card? It’s the full desktop die with 448 GB/s of GDDR7, so inference speed matches a tower 5060 Ti. The open question is sustained thermals in 2.65 liters on a 330W budget; early coverage is gaming-focused.

Why not the Magnus One with the RTX 5070 Ti for AI? Same 16GB VRAM wall at nearly twice the price (~$3,000 barebone). Its 896 GB/s makes already-fast small models faster, but it unlocks zero additional models. For $3,000+ you’re in EVO-X2 or used-dual-GPU territory, both of which run things the 5070 Ti can’t.

Can the EN275060TC’s system RAM help run bigger models? You can offload layers to its DDR5 (up to 96GB), but SODIMM bandwidth turns decode into single digits — the same 2–5 tok/s penalty we measure on every CPU-offload setup. Treat 16GB as the real ceiling for usable speed.

Should I wait for the EAMAX395C anyway? Only if Zotac’s price lands meaningfully under $3,400 — and nothing suggests it will, given 128GB of LPDDR5X now costs OEMs more than the APU does. Identical performance is on shelves today.

Sources

Last updated October 2, 2026. Prices and specs change — DRAM and GPU pricing have both moved double digits within single months this year; verify current rates before purchasing.

Was this article helpful?

Get the numbers before you buy

New GPU and mini-PC benchmarks, VRAM thresholds, and price checks — sent when there's something worth acting on, not on a schedule. No spam, unsubscribe anytime.