TL;DR: 128GB unified memory buys the 100B–276B MoE class that no consumer graphics card holds — gpt-oss-120b at 34–56 tok/s, Inkling-Small 276B at 2-bit, Qwen3-235B at ~11 tok/s. The catch: only ~96–120GB is GPU-usable, and 256–273 GB/s bandwidth makes dense 70Bs crawl. Buy this tier for big sparse MoE models or don’t buy it at all.

	Strix Halo box (Framework, EVO-X2)	NVIDIA DGX Spark	Mac Studio M4 Max 128GB
Best for	Cheapest 128GB, Linux tinkerers	CUDA stack, concurrent serving	Fastest single-stream at this price
Price (Aug 2026)	$3,449–$3,649	$4,699	~$3,699
Bandwidth	256 GB/s	273 GB/s	546 GB/s
The catch	96GB Windows VGM cap (Linux fixes it)	Was $3,999 — repriced +$700	macOS wires only ~96GB to GPU by default

Honest take: the Mac Studio M4 Max is the best all-around 128GB machine — double the bandwidth of the x86 boxes at the same street price. Buy Strix Halo only if you need Linux/CUDA-free economics, and DGX Spark only if you need the CUDA toolchain in a small box.

Our VRAM-tier series ended at 96GB because discrete cards end there — above it, single-GPU means the H200 class NVIDIA doesn’t sell at retail. But there’s a fourth path the series hasn’t given its own page: 128GB of unified memory, shared between CPU and GPU. Three very different machines sell it in 2026 — AMD’s Strix Halo mini-PCs, NVIDIA’s DGX Spark, and Apple’s 128GB M4 Max / M5 Max — and they all run the same question: which models actually make sense on lots of slow memory?

The answer is narrower than the spec sheet suggests, and it starts with a number that isn’t 128.

The first trap: 128GB is not 128GB

Every machine in this tier reserves a chunk of the pool for the OS, and each does it differently. This is the single most common reason a model that “should fit” refuses to load.

Windows on Strix Halo: AMD’s Adrenalin software (or the BIOS) lets you dedicate at most 96GB of a 128GB system as graphics memory, per jdhodges’ allocation guide. Anything larger than ~96GB minus KV cache is out, full stop. Worse, LM Studio’s ROCm runtime historically failed to see shared memory at all when the BIOS was set to Auto (lms issue #589).

Linux on Strix Halo: leave the BIOS allocation at Auto/512MB and raise the kernel’s GTT limits instead. Zypher Systems’ guide documents pushing GPU-accessible memory to ~120GB:

# /etc/default/grub — append to GRUB_CMDLINE_LINUX_DEFAULT
ttm.pages_limit=31457280 amdgpu.gttsize=120000

After update-grub and a reboot, llama.cpp (Vulkan or ROCm) can map ~120GB on demand — this is what makes the 101GB Qwen3-235B quant possible on this hardware at all. AMD’s own ROCm Strix Halo optimization docs recommend the same TTM-parameter approach.

macOS: the Metal driver wires roughly 75% of RAM to the GPU by default — 96GB on a 128GB Mac, per ModelPiper’s wired-limit writeup. One sysctl raises it:

sudo sysctl iogpu.wired_limit_mb=122880   # ~120GB; resets on reboot

Leave ~8GB for macOS or the machine starts paging mid-generation.

DGX Spark: NVIDIA’s software stack keeps its own reserve; plan for ~120GB usable of the 128GB LPDDR5X.

So the real planning number for this tier is 96GB out of the box, ~120GB tuned. Every fit claim below uses that.

The second trap: bandwidth decides everything else

Decode speed on local LLMs is memory-bandwidth-bound — the GPU re-reads every active weight for each token. The x86 boxes ship 256–273 GB/s; that’s less than a third of a used RTX 3090’s 936 GB/s. A dense 70B at Q4 (~43GB of weights per token) lands around 4–5 tok/s on Strix Halo and, measured by CraftRigs, 15–22 tok/s on the M4 Max’s 546 GB/s. Fine for overnight batch work, painful for chat.

Sparse MoE models flip the math: a 120B MoE with ~5B active parameters reads ~3–4GB per token instead of 43GB. That’s why every model worth buying this tier for is a Mixture-of-Experts — you’re paying for capacity to hold the experts, and the small active set keeps decode speed civilized. Our Why Local LLMs Got Good deep dive covers the mechanics.

What to run: the 128GB verdict table

Model	Quant / size	Fits at 96GB default?	Measured speed	Best at
gpt-oss-120b (117B-A5B)	MXFP4, 63.4GB	Yes, with big KV room	34–56 tok/s Strix Halo, 38–52 DGX Spark	Default pick: reasoning + tools
Nemotron 3 Super (120B-A12B)	UD-IQ4_XS, ~60–65GB	Yes	18.4 tok/s Strix Halo	Agentic coding, long context
Inkling-Small (276B-A12B)	UD-Q2_K_XL, 87.9GB	Yes, barely (~8GB KV)	~20–30 tok/s Strix Halo (estimate)	Frontier-class quality ceiling
Qwen3-235B-A22B	Q3, ~101GB	No — needs Linux GTT / Mac sysctl	~11 tok/s Strix Halo	Maximum knowledge depth
Llama 4 Scout (109B-A17B)	Q4_K_M, ~55GB	Yes, easily	Usable; bandwidth-bound	Long-context multimodal

gpt-oss-120b is the reason this tier exists. OpenAI ships it MXFP4-native at 63.4GB, and with only ~5B parameters active per token it’s the fastest big model on slow memory: community llama.cpp benchmarks put Strix Halo at up to 55.57 tok/s tg128 on Vulkan/RADV, with Hardware Corner’s head-to-head measuring 34.13 tok/s (Strix Halo) vs 38.55 tok/s (DGX Spark) on identical settings — backend and llama.cpp version explain the spread. Either way it’s reading-speed-plus, with ~30GB left over for KV cache at long context. Our gpt-oss 20B guide covers the small sibling for 16GB cards.

Nemotron 3 Super 120B-A12B is the agentic-coding pick. The Strix Halo wiki’s llama.cpp table measured the 4-bit UD-IQ4_XS build at 18.43 tok/s tg128 — slower than gpt-oss-120b because 12B active parameters is 2.4× the memory traffic of 5B. Still above the ~7–10 tok/s reading-speed floor.

Inkling-Small 276B is the quality ceiling. Thinking Machines’ Unsloth 2-bit dynamic GGUF is 87.9GB — the only frontier-adjacent checkpoint that squeezes under the 96GB default line, with roughly 8GB left for KV (plan on 8–16K context, not the headline 1M). The 12B active set keeps decode in the same ballpark as Nemotron Super; treat our ~20–30 tok/s Strix Halo figure as an estimate scaled from measured 12B-active models, not a benchmark. Full details in our Inkling-Small hardware guide. The 2-bit quality tax remains unmeasured — test before you trust it for real work.

Qwen3-235B-A22B is what the tuned ~120GB ceiling is for. The Q3 build runs about 101GB and won’t load under Windows’ 96GB VGM cap — this is a Linux-GTT or raised-Mac-limit model only. Real-world speed is ~11 tok/s on Strix Halo: usable for chat, generous in knowledge depth, and the single biggest model you can hold on one sub-$4,000 machine.

Llama 4 Scout fits with 40GB to spare — Unsloth’s Q4_K_M lands around 55GB — and its 17B active parameters put decode between the Qwen and Nemotron numbers. Take it for the 10M-token context marketing only after reading our Scout VRAM reality check.

And if your daily driver is a 30B-class MoE — GLM-4.7-Flash 30B-A3B, Qwen3.6-35B-A3B — a 128GB box runs them at 70–100+ tok/s while holding a second big model resident. But if that’s all you run, stop: a 24GB used RTX 3090 at $1,201–$1,299 (ResalePrices, Aug 2026 — $1,264 average across 366 listings) does the same job 2–3× faster for a third of the money.

What you CANNOT run at 128GB

The 2026 flagship MoEs jumped clean over this tier:

Model	Smallest useful quant	Why it misses
GLM-5.2 (744B)	UD-IQ2_M, 239GB	Needs 256GB unified or a 4× RTX PRO 6000 build
Kimi K2.7/K3 (1T)	2-bit, ~245–325GB	2–3× over budget even tuned
MiniMax M3 (428B)	Q2_K_XL, 143GB	~23GB over the tuned 120GB ceiling
Qwen3-Coder 480B	UD-IQ1_M, ~150GB	1-bit and still doesn’t fit
Dense 70B at speed	Fits fine at Q4 (43GB)	4–5 tok/s on 256 GB/s — capacity isn’t the problem, bandwidth is

That last row is the honest one: 128GB holds a dense 70B in its sleep, but on the x86 boxes it decodes below reading speed. If dense 70Bs are your target, a 48GB dual-3090 rig beats every machine on this page.

The machines, priced (August 2026)

The DRAM crisis hit this tier hardest — every price below is up 40–75% from launch:

Machine	Aug 2026 price	Launch price	Bandwidth
Framework Desktop 128GB	$3,449 (pre-order)	$1,999	256 GB/s
GMKtec EVO-X2 128GB	$3,499 (1TB) / $3,649 (2TB)	$1,999–$2,199	256 GB/s
NVIDIA DGX Spark	$4,699	$3,999	273 GB/s
Mac Studio M4 Max 128GB	~$3,699	—	546 GB/s
MacBook Pro 16” M5 Max 128GB	$6,699+ as configured	—	614 GB/s

Pricing sources: ComputingForGeeks’ Strix Halo comparison (Framework $3,449, up from $1,999), VideoCardz (EVO-X2), NVIDIA’s February price-change notice via OC3D (Spark +$700 on “memory supply”), LLMPicker (Mac Studio), Micro Center (M5 Max 128GB, see Sources).

The uncomfortable result: Apple — historically the RAM-upcharge villain — now sells the fastest 128GB machine at the same street price as the x86 boxes, because Apple’s prices held while DRAM-crisis repricing ate AMD’s value story. The M4 Max’s 546 GB/s doubles every decode number in this article versus Strix Halo; our M4 Max vs RTX 5090 comparison has the measured ladder, and the M5 Max adds another ~12% bandwidth (614 GB/s, ~28 tok/s on 70B Q4 per LLM Check) if you need it in a laptop.

Pick DGX Spark only when the CUDA stack is the point — fine-tuning, TensorRT, vLLM concurrency (where it aggregates 862 tok/s on gpt-oss-120b at 256 streams, a use case none of the others touch). For single-user chat it’s the worst $/tok/s on this page.

If you want the next tier up

Above 128GB, the ladder gets steep. The 239GB GLM-5.2 quant needs 192–256GB: that’s a second RTX PRO 6000 96GB stacked (~$32,000 after the August repricing), a 4× used RTX 3090 box with 192GB+ system RAM running hybrid offload, a 2-node Strix Halo cluster over llama.cpp RPC, or a Mac Studio RDMA cluster. All of them cost more than renting: an H200 pod on RunPod by the hour is the sane way to find out whether you actually need 400B-class local inference before committing five figures. Sanity-check any of these against the GPU buying guide and the VRAM calculator first.

FAQ

Is 128GB unified memory better than a 24GB GPU? Different tools. The 3090 is 2–3× faster on everything that fits in 24GB; the 128GB box runs 100B–276B MoE models the 3090 can’t hold at any speed. If you don’t have a specific big-MoE use case, the 24GB tier is the better buy at a third of the price.

Why does my 100GB model fail to load on a 128GB machine? The GPU can’t touch all 128GB. Windows Strix Halo caps at 96GB VGM; macOS wires ~96GB by default. Fix: Linux GTT kernel parameters (ttm.pages_limit + amdgpu.gttsize) on Strix Halo, sudo sysctl iogpu.wired_limit_mb=122880 on macOS.

Should I wait for cheaper 128GB machines? The trend runs the other way — Framework went $1,999 → $3,449, Spark $3,999 → $4,699, and DRAM contract prices are still climbing. If the workload justifies it, buying now beats waiting; if it doesn’t, rent.

What about running these models on the NPU? Don’t. On every 2026 machine we’ve tested or tracked, LLM decode runs on the GPU (Vulkan/ROCm/Metal); the NPU path is slower or unsupported — see NPU vs GPU and the Strix Halo VGM trap.

Recommended Gear

• GMKtec EVO-X2 128GB — Strix Halo in stock form, $3,499–$3,649
• Mac Studio M4 Max 128GB — the bandwidth pick at ~$3,699
• RTX 3090 — still the speed-per-dollar king if 24GB is enough
• RTX PRO 6000 Blackwell — the 96GB discrete alternative
• MacBook Pro 16” M5 Max — the fastest 128GB option, in laptop form at $6,699+

For local coding stacks that point Cursor or Cline at one of these boxes, our sister site aicoderscope.com covers the tooling side.

Sources

• AMD Strix Halo backend benchmarks (gpt-oss-120b 55.57 tok/s tg128) — kyuz0
• First DGX Spark LLM benchmarks vs Strix Halo (34.13 vs 38.55 tok/s) — Hardware Corner
• llama.cpp performance table (Nemotron 3 Super 18.43 tok/s) — Strix Halo Wiki
• gpt-oss: how to run (MXFP4 63.4GB) — Unsloth Docs
• Inkling: how to run locally (Inkling-Small UD-Q2_K_XL 87.9GB) — Unsloth Docs
• Llama 4: how to run (Scout Q4_K_M ~55GB) — Unsloth Docs
• Is a Strix Halo mini PC worth buying for local AI? (Qwen3-235B Q3 ~101GB, ~11 tok/s) — Popular AI
• Pushing Strix Halo to 120GB unified VRAM under Linux — Zypher Systems
• Strix Halo system optimization — AMD ROCm Documentation
• How to allocate VRAM on Strix Halo (96GB Windows cap) — jdhodges
• iogpu.wired_limit_mb: raising the Metal memory ceiling — ModelPiper
• Ryzen AI Max+ 395 mini PCs compared: real prices — ComputingForGeeks
• GMKtec EVO-X2 with 128GB debuts at $3,500 — VideoCardz
• NVIDIA raises DGX Spark price by $700 — OC3D
• DGX Spark benchmarks: concurrency (862 tok/s aggregate) — Tokenstead
• Mac Studio M4 Max 128GB: 70B at 15–22 tok/s tested — CraftRigs
• M5 Max for local AI (614 GB/s, ~28 tok/s 70B Q4) — LLM Check
• MacBook Pro 16” M5 Max 128GB listing — Micro Center
• RTX 3090 used price tracker ($1,264 avg, Aug 2026) — ResalePrices
• How to run GLM-5.2 locally (UD-IQ2_M 239GB) — Thunder Compute
• Two-node Strix Halo cluster with llama.cpp RPC — Framework Community

Last updated August 22, 2026. Prices and specs change; verify current rates before purchasing.