TL;DR: Every Ryzen AI Max+ 395 mini PC solders its LPDDR5X, so the 64GB-or-128GB choice is permanent. 64GB caps your GPU pool at ~48GB — enough for 70B dense and 80B MoE at Q4, never for gpt-oss-120b. If the 100B+ class is why you want this platform, only 128GB delivers it, and the gap is $1,250–$1,490.

	64GB Strix Halo	128GB Strix Halo	Used RTX 3090 tower
Best for	70B/80B Q4 ceiling on a budget	120B MoE, long context, multi-model	Everything ≤32B, much faster
Price / Cost	$1,959–$2,199	$3,449–$3,649	~$1,150–$1,350 (card)
The catch	No gpt-oss-120b, ever — RAM is soldered	120B tops out at 31–56 tok/s; volatile pricing	24GB wall; 70B needs a second card

Honest take: If Strix Halo makes sense for you at all, buy 128GB — the models that justify this platform’s existence don’t fit in 64GB, and you cannot add RAM later. If $3,400+ is out of reach, a used RTX 3090 beats the 64GB config at everything that fits in 24GB, for less money.

Strix Halo machines — the GMKtec EVO-X2, Framework Desktop, Beelink GTR9 Pro, Minisforum MS-S1 Max — all ship the same AMD Ryzen AI Max+ 395 silicon: 16 Zen 5 cores, a 40 CU Radeon 8060S iGPU, and 256 GB/s of LPDDR5X bandwidth. We’ve compared the boxes against each other, against a Mac Studio, and against a DGX Spark. What no page covered is the decision that comes before picking a box: which memory configuration.

It matters more here than on any tower build, for one blunt reason: the LPDDR5X is soldered to the board on every one of these machines. There is no SO-DIMM slot. The capacity you order is the capacity that machine dies with. And in September 2026, the price step between the two tiers is $1,250–$1,490 — real money that buys you exactly one thing: the 100B+ model class.

Check your target model against your candidate config in our VRAM calculator before you order. Here’s the full math.

What you can actually allocate (it’s not 64 or 128)

Strix Halo has no dedicated VRAM. The GPU and CPU share one LPDDR5X pool, and how much of it the GPU can claim depends on the OS:

On Windows, AMD’s Variable Graphics Memory (Adrenalin → Performance → Tuning) converts up to 75% of total RAM into “dedicated” graphics memory (AMD’s own FAQ, driver 25.6.1+). That means:

• 64GB machine → 48GB GPU pool max
• 128GB machine → 96GB GPU pool max

On Linux, you keep the BIOS carve-out minimal and raise the GTT (Graphics Translation Table) limit with kernel parameters instead — the same fix we documented in the Framework Desktop comparison. On a 128GB machine:

# /etc/default/grub — 128GB machine, ~110GB GPU-addressable
GRUB_CMDLINE_LINUX_DEFAULT="ttm.pages_limit=31457280 amdgpu.gttsize=120000"
$ sudo update-grub && sudo reboot
$ sudo dmesg | grep -i "amdgpu.*gtt"
[    4.312] [drm] amdgpu: 120000M of GTT memory ready.

On a 64GB machine the same formula scaled down (amdgpu.gttsize=57344, leaving ~8GB for the OS) gets you to roughly 56GB — a few GB past the Windows cap, but nowhere near enough to change which model class fits. Jeff Geerling’s APU guide and the official ROCm Strix Halo docs cover the mechanics.

So the real comparison is a ~48GB GPU pool vs a ~96–110GB GPU pool. Write those numbers on a sticky note; every line below follows from them.

What each pool actually runs

Model weights are only part of the budget — KV cache grows with context, and the OS keeps several GB for itself. Verified GGUF sizes against both pools, with measured Strix Halo speeds:

Model	Quant	Weights	64GB (48GB pool)	128GB (96GB+ pool)	Measured speed
Qwen3-30B-A3B (MoE)	IQ4_XS	~18 GB	✅ comfortable	✅ trivial	100 tok/s
Qwen3.6-35B-A3B (MoE)	Q4_K_M	~20 GB	✅ comfortable	✅ trivial	~77 tok/s
Llama 3.3 70B (dense)	Q4_K_M	~40–43 GB	✅ tight, short context	✅ with long context	~4.7–5 tok/s
Qwen3-Next-80B-A3B (MoE)	UD-Q4_K_XL	~46 GB	⚠️ barely, tiny context	✅ full 262K context	~59–63 tok/s
gpt-oss-120b (MoE)	MXFP4	~61 GB (~65GB resident)	❌ does not fit	✅ comfortable	31–56 tok/s (see below)
Two models resident (agent + coder)	—	60GB+	❌	✅	—

Speeds: the 30B-A3B figure is the Strix Halo benchmark canon (llama.cpp, RADV Vulkan); the 35B-A3B, 70B, and 80B-A3B rows come from the cross-OEM strix-halo-guide benchmark set (llama-bench, Vulkan RADV, re-run September 26, 2026). On gpt-oss-120b the spread is real: ServeTheHome measured ~31 tok/s at 120–128W on Ollama defaults, while tuned llama.cpp Vulkan RADV runs hit 55–56 tok/s (strix-halo-guide, See Hiong’s build log). Plan around the low end; treat the high end as what tuning buys you.

Three things jump out of that table.

First: the 64GB config is genuinely capable. A 48GB pool holds Llama 3.3 70B Q4_K_M — the thing that made this platform famous, since no single consumer discrete GPU can — and Qwen3-Next-80B-A3B at ~59 tok/s, which is arguably the best daily-driver model on the whole platform. If your ambitions end at the 70B/80B Q4 tier, 64GB does the job for $1,250–$1,490 less.

Second: the ceiling is hard and close. gpt-oss-120b’s MXFP4 weights alone (~61GB on disk, ~65GB resident per MindStudio’s measurements) blow past a 48GB pool before the KV cache writes its first byte. No quant of the 100B+ class squeezes in. And the 80B-A3B that does fit has almost no room left: ~46GB of weights in a 48GB pool leaves ~2GB for KV cache — on Windows it’s effectively a benchmark result, not a workflow; the Linux GTT route to ~56GB is what makes it usable, and even then you’re running a fraction of the 262K context the model supports.

Third: capacity is the only thing you’re buying. Both configs have identical 256 GB/s bandwidth, so every model that fits in 48GB runs at the same speed on both machines. The extra $1,400 buys zero tokens per second on anything the 64GB box can load. It buys the models the 64GB box can’t load at all.

What it looks like when you guess wrong

Buy the 64GB config, then try the model everyone bought this platform for:

$ llama-server -m gpt-oss-120b-MXFP4.gguf -ngl 99 --ctx-size 8192
ggml_vulkan: Device memory allocation of size 4294967296 failed.
ggml_gallocr_reserve_n: failed to allocate Vulkan0 buffer of size 65498447872
llama_model_load: error loading model: unable to allocate backend buffer

There’s no flag that fixes this. Your options, in descending order of sanity: run a smaller model (Qwen3-Next-80B-A3B is genuinely close in quality per public benchmarks), offload experts to “CPU” memory — which on a unified-memory machine is the same 64GB pool you already exhausted — or rent the capacity: a cloud instance on Vast.ai (RTX 3090 from ~$0.07/hr, market pricing) lets you verify a model’s real usefulness before you’ve spent workstation money on it. What you cannot do is add RAM. That’s the entire argument for deciding this correctly the first time.

The price math, September 2026

Verified street prices this week:

Machine	64GB	128GB	Step
Framework Desktop (DIY, no SSD/OS)	$1,959	$3,449	+$1,490
GMKtec EVO-X2 (2TB, Win11)	$1,999–$2,199	$3,499–$3,649	+$1,300–$1,450

Sources: Framework configurator via ComputingForGeeks’ September re-check, Amazon’s EVO-X2 64GB listing, GMKtec direct, and PriceHistory tracking on the 128GB SKU — which, fair warning, swung from $2,199 to $3,649 in the single week of September 20–26. Strix Halo pricing is the most volatile of anything we track; treat every number here as a snapshot.

That step works out to roughly $20–23 per GB of soldered LPDDR5X. Eighteen months ago that would have been an outrageous markup. In the current DRAM crisis — where a plain DDR5 64GB kit runs $680–$1,070 ($10.60–$16.70/GB) and can’t feed a GPU at 256 GB/s anyway — the premium is uncomfortably reasonable. The RAM-crisis rule we keep repeating (“buy what the workload needs, then stop”) cuts both ways here: don’t pay $1,400 for capacity you’ll never load, but don’t save $1,400 into a ceiling you’ll hit in month two either.

Which one you actually need

Be honest about which row you’re in:

You want a local coding assistant or chat on 8B–32B models. Wrong platform entirely. A used RTX 3090 at $1,150–$1,350 has 3.7× the memory bandwidth (936 vs 256 GB/s), which is the number that sets decode speed on dense models — 161 tok/s on gpt-oss-20b against Strix Halo’s MoE-assisted ~100 tok/s ceiling, and a far larger gap on prompt processing. The 3090 remains the value king for everything that fits in 24GB.

Your ceiling is 70B dense or 80B MoE at Q4, and budget is tight. The 64GB config is the cheapest hardware on Earth that holds a 70B Q4 model entirely in GPU-addressable memory. Accept the ~4K context limit on 80B-A3B and the ~5 tok/s reality of dense 70B, and you’ve saved ~$1,400.

You want gpt-oss-120b, long context on big MoE, or an always-loaded pair of models. 128GB, no debate. This is the configuration the platform was designed around, and per our 128GB model guide it’s the cheapest path to the 100B+ class that exists — a quad-3090 build with the same capacity starts at $5,600 and pulls 1,400W.

You’re not sure your workload justifies any of this. Rent first. An hour of Vast.ai time against the exact model you’re considering costs less than a coffee and answers the question with data instead of a forum thread.

What to actually buy

Prices as of September 2026, all verified above:

Your situation	The machine	Price	Where
Everything you run fits in 24GB	Used RTX 3090	~$1,150–$1,350	Check price
70B/80B Q4 is the ceiling, budget is tight	GMKtec EVO-X2 64GB	~$1,999–$2,199	Check price
120B-class models, long context — the platform’s whole point	Framework Desktop 128GB	~$3,449	frame.work
Same, but you want it prebuilt with SSD + Windows	GMKtec EVO-X2 128GB	~$3,499–$3,649	Check price
Undecided — test the workload first	Rented GPU, from $0.07/hr	pay per hour	Vast.ai

FAQ

Can I upgrade a 64GB Strix Halo machine to 128GB later? No. The LPDDR5X is soldered on every Ryzen AI Max+ 395 machine shipping in 2026 — EVO-X2, Framework Desktop, GTR9 Pro, MS-S1 Max, all of them. Signal integrity at 8000 MT/s across a 256-bit bus is the stated reason. The config you buy is final.

Is the 128GB version faster? No. Same chip, same 256 GB/s bandwidth, same tok/s on any model that fits both. You’re buying capacity, not speed.

Does the 32GB config make sense for local AI? Skip it. A 32GB machine caps the GPU pool around 24GB — the same ceiling as a used RTX 3090 that’s roughly half the price and 2–4× faster.

What about the 96GB configs some vendors list? Rare in practice (a few laptop SKUs and the Beelink GTR9 Pro’s mid tier). A 96GB machine gets a ~72GB Windows VGM pool — enough for gpt-oss-120b, tight for anything bigger. If the price lands meaningfully below the 128GB tier it’s a defensible middle; usually it doesn’t.

Should I wait for the Ryzen AI Max+ P495 (Gorgon Halo) refresh? Our standing verdict is no: it’s the same silicon with ~7% more memory bandwidth (LPDDR5X-8533 vs 8000) and no confirmed pricing. The 64-vs-128 decision will be identical on it.

Recommended Gear

• GMKtec EVO-X2 (Ryzen AI Max+ 395) — 64GB for the 70B/80B ceiling, 128GB for the 120B class
• Used RTX 3090 24GB — faster than either config for everything ≤32B

For the software side of a Strix Halo box — ROCm vs Vulkan backends, Ollama setup, GGUF sourcing — see our sister site’s guides at aifoss.dev, and if the machine will back a local coding assistant, aicoderscope.com covers wiring it into Cline and Continue.

Sources

• FAQs: AMD Variable Graphics Memory, VRAM, AI Model Sizes — AMD
• AMD Strix Halo system optimization — ROCm Documentation
• Increasing the VRAM allocation on AMD AI APUs under Linux — Jeff Geerling
• How to Allocate VRAM on Strix Halo: AMD Adrenalin Guide — jdhodges
• Beelink GTR9 Pro AMD Ryzen AI Max+ 395 Review — ServeTheHome
• strix-halo-guide: cross-OEM llama.cpp benchmarks — GitHub
• Running Llama.cpp on AMD Strix Halo — See Hiong
• AMD Ryzen AI Max+ 395: 128GB Unified Memory for Local LLMs — MindStudio
• Qwen3-Next-80B-A3B-Instruct GGUF — Unsloth on Hugging Face
• GMKtec EVO-X2 64GB listing — Amazon
• GMKtec EVO-X2 128GB price history — PriceHistory
• Ryzen AI Max+ 395 Mini PCs Compared: Real Prices — ComputingForGeeks

Last updated September 28, 2026. Prices and specs change — Strix Halo pricing has moved by four figures within a single week this month. Verify current rates before purchasing.

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.

Email address Subscribe