DGX Spark vs Strix Halo vs Mac Studio: which local-AI box to buy

For the first time, a 70B-parameter model runs on a single quiet desktop instead of a multi-GPU tower -- and in 2026 three very different machines get you there. NVIDIA's GB10 "DGX Spark" boxes bring CUDA and a dedicated FP4 pipeline; AMD's Strix Halo mini-PCs put 128 GB of unified memory in reach at a lower cost than the rivals; and Apple's Mac Studio answers models the fastest of the three. They share the trick that makes this possible -- unified memory, where the CPU and GPU address one large pool -- but they are not interchangeable, and the marketing numbers won't tell you which to buy.

The honest tie-breaker is a single spec: memory bandwidth. Generating text is memory-bandwidth-bound, so bandwidth (GB/s) is the primary driver of tokens per second, while unified-memory capacity decides which models fit at all. Almost every competing guide leads with NPU "TOPS" instead -- a number that barely touches large-model speed. This comparison leads with bandwidth, backs it with measured token rates, and ends with a clear pick by use case.

This comparison was produced with AI assistance as part of VerdictBits's research workflow. Every spec, price, and Amazon listing was cross-checked against primary sources before publication; token-per-second figures come from independent testers using different quantisations and backends, so we cite them as directional, not a single controlled benchmark, and we flag prices as volatile -- re-check before you buy.

The verdict, at a glance

Why memory bandwidth decides it -- not TOPS

Every token a language model generates requires the machine to read its active weights out of memory and do a relatively small amount of arithmetic on them. The arithmetic is cheap; the memory read is the bottleneck. So the number that predicts token speed is memory bandwidth, not the NPU's "TOPS" rating -- which measures raw arithmetic the mainstream runtimes (llama.cpp, Ollama, MLX) don't even use for this, because they run on the GPU. Capacity decides which models load; bandwidth decides how fast they answer.

Strix Halo (Ryzen AI Max+ 395) real; 256 theoretical 215 GB/s

DGX Spark / GB10 273 GB/s

Mac Studio M4 Max 546 GB/s

Mac Studio M3 Ultra 512 GB tier, Apple-direct 819 GB/s

Higher is faster token generation. Strix Halo figure is real measured GPU bandwidth (rocm_bandwidth_test); others are package specs. Sources: Level1Techs, Framework community, NVIDIA, Apple.

That single chart explains most of the head-to-head. The Mac Studio M4 Max moves weights about twice as fast as the GB10 or Strix Halo boxes, so it generates tokens roughly twice as fast on models all three can hold. The Strix Halo box holds the most memory per dollar but streams it the slowest; the GB10 sits in the middle on bandwidth but wins on software and prompt-processing. Everything below follows from those trade-offs.

Head-to-head specs

Token-per-second figures are directional -- independent testers use different quantisations, context lengths, and backends (llama.cpp / Ollama / vLLM / MLX), so treat them as ballpark, not a controlled chart. MoE = mixture-of-experts.

The three platforms

NVIDIA DGX Spark / GB10 -- The CUDA box

GB10 Grace-Blackwell - 20-core Arm + Blackwell GPU with FP4 · 128 GB LPDDR5X unified · 273 GB/s · CUDA + DGX OS

Verdict. The only CUDA + FP4 box here, with the best prompt-processing of the three - but its 273 GB/s makes token generation no faster than a Mac mini that costs half as much. Buy it for the NVIDIA software stack, not raw decode speed.

Reasons to buy

• Full CUDA / DGX OS ecosystem - the widest tooling of the three
• Dedicated FP4 hardware and excellent prefill (prompt processing)
• 128 GB unified holds a 70B dense or ~200B MoE model at 4-bit
• Two units link over dual 200 GbE (ConnectX-7) to hold roughly 405B

Reasons to avoid

• 273 GB/s caps decode - only ~3-5 tok/s on a dense 70B model
• Most expensive entry (GX10 ~$3k+, DGX Spark FE $4,699)
• Price already rose ~18% on LPDDR5X shortages (Feb 2026)
• Overkill unless you specifically need the NVIDIA stack

Best for: CUDA tooling, fast prompt-processing, prototyping the NVIDIA stack.

Buy it as: ASUS Ascent GX10 — from ~$3,000 (1 TB); DGX Spark Founders Edition $4,699

the on-Amazon GB10 box - verify the storage tier before buying, price swings a lot by config

Check current price on Amazon →

AMD Strix Halo (Ryzen AI Max+ 395) -- The value box

Ryzen AI Max+ 395 - Radeon 8060S (40 CU) + XDNA 2 NPU · up to 128 GB LPDDR5X-8000 unified (~96 GB addressable as GPU) · ~215 GB/s real (256 GB/s theoretical) · ROCm / Vulkan / llama.cpp

Verdict. The lower-cost route into 128 GB of unified memory - it loads models a 24 GB GPU cannot touch, and lower configs are genuinely cheap. But the LPDDR5X shortage pushed a loaded 128 GB / 2 TB box to ~$4,349 (near the DGX Spark), so it is no longer a runaway bargain. And its lowest real bandwidth of the three (~215 GB/s) plus the least mature software mean it shines on MoE models and on 'fits at all', not on fast dense-70B speed.

Reasons to buy

• The lower-cost 128 GB platform, especially at promo / lower-SSD configs
• Runs GPT-OSS 120B (MoE) at ~34 tok/s - punches above its price
• Low power draw, XDNA NPU, and Amazon-available
• Up to ~96 GB of the 128 GB addressable as GPU memory

Reasons to avoid

• Lowest real bandwidth (~215 GB/s) - a dense 70B is single-digit tok/s
• ROCm / Vulkan still maturing next to CUDA and MLX
• The LPDDR5X shortage pushed loaded 128 GB units to ~$4,349
• A 120 W BIOS cap limits external-GPU (eGPU) expansion

Best for: the lower-cost route into 128 GB unified; big mixture-of-experts models without a data-center GPU.

Buy it as: Beelink GTR9 Pro — ~$2,000-4,349 (loaded 128 GB / 2 TB ~$4,349)

the LPDDR5X shortage pushed a loaded 128 GB / 2 TB unit to ~$4,349; lower-memory or promo configs run less - check the exact config

Check current price on Amazon →

Apple Mac Studio (M4 Max) -- The bandwidth box

Apple M4 Max, up to 40-core GPU (M3 Ultra for the 512 GB tier) · up to 128 GB (M4 Max) / 512 GB (M3 Ultra) unified · 546 GB/s (M4 Max); 819 GB/s (M3 Ultra) · MLX / Metal

Verdict. The best token-generation speed per dollar in the middle of the market: the M4 Max's 546 GB/s is roughly double the GB10 and Strix Halo boxes, so it answers a 70B model markedly faster while staying near-silent. You pay Apple prices and lose CUDA, and huge dense models become GPU-compute-limited - but for most local-LLM buyers it is the most balanced box.

Reasons to buy

• 546 GB/s (M4 Max) - about 2x the GB10 / Strix decode bandwidth
• MLX ecosystem is polished and improving quickly
• Near-silent, low-watt, and genuinely turnkey
• M3 Ultra scales to 512 GB at 819 GB/s for giant MoE models

Reasons to avoid

• No CUDA - some tooling is still NVIDIA-first
• Big dense models become GPU-compute-limited, not RAM-limited
• The 512 GB M3 Ultra is ~$9,500 - steep at the very top
• Not sold at every config on Amazon (buy Apple / Best Buy)

Best for: the fastest token generation per dollar in the middle of the market; quiet and turnkey.

Buy it as: Apple Mac Studio (M4 Max) — from ~$1,999 (M4 Max); M3 Ultra 512 GB ~$9,500

roughly 2x the mini-boxes' bandwidth; near-silent and low-watt. Sold via Apple / Best Buy too

Check current price on Amazon →

Which to buy, by use case

Want the specific box within each platform -- exact SKUs, storage tiers, and the cheapest current listing -- rather than the platform-level call? Our best AI PC for running LLMs locally roundup ranks the individual machines, and the LLM VRAM calculator tells you exactly how much unified memory your target model needs.

Gotchas before you buy

• Ignore the "TOPS" number. A box advertised at "126 TOPS" doesn't generate text faster -- that figure is the NPU's arithmetic rating, and llama.cpp / Ollama / MLX run on the GPU. Judge by bandwidth and memory capacity instead.
• MoE vs dense changes everything. A cheap Strix Halo box runs a 120B mixture-of-experts model faster than a dense 70B, because MoE activates only a few experts per token. Match the benchmark you read to the model architecture you actually plan to run.
• Prices are volatile and buy-boxes are inflated. The DGX Spark already rose ~18% ($3,999 to $4,699) on memory shortages, and third-party sellers mark up the Strix Halo minis by hundreds on Amazon. Check the seller and compare against the vendor's own price before buying.
• Software maturity is real. CUDA (NVIDIA) and MLX (Apple) are smoother today than AMD's ROCm / Vulkan path. If you don't want to troubleshoot backends, weight that in.
• The Strix Halo eGPU cap. These boxes cap external-GPU power delivery around 120 W, so don't plan on bolting on a big discrete card to fix the bandwidth ceiling -- it won't.

Frequently asked questions

Which is fastest on a 70B model - DGX Spark, Strix Halo, or Mac Studio?

The Mac Studio, and it is not close in the middle tier. Token generation is memory-bandwidth-bound, and the M4 Max streams weights at 546 GB/s - roughly double the DGX Spark / GB10 (273 GB/s) and the Strix Halo box (~215 GB/s real). On a dense 70B at 4-bit the two mini-boxes land in the single digits to low teens of tokens per second, while an M4 Max pulls clearly ahead, and Apple's 819 GB/s M3 Ultra further still. All three can hold the model; the Mac answers it faster.

Is the DGX Spark worth it over a Mac Studio for local LLMs?

Only if you specifically need CUDA. The GB10 has a real edge in prompt-processing (prefill) thanks to its FP4 hardware, and it runs the full NVIDIA / DGX software stack that a lot of AI tooling assumes. But for pure token-generation speed its 273 GB/s bandwidth is no faster than a Mac mini M4 Pro that costs less than half as much, and the Founders Edition is $4,699. If your work lives in CUDA, buy it; if you just want fast local inference, a Mac Studio M4 Max is the better value.

Can AMD Strix Halo really run a 120B model?

Yes - if it is a mixture-of-experts (MoE) model. A Ryzen AI Max+ 395 box with 128 GB unified runs GPT-OSS 120B at roughly 34 tokens/second in independent tests, because MoE models activate only a few experts per token, so the effective memory read per token is far smaller than the full 120B. A dense 70B on the same box is much slower (single digits to ~14 tok/s) precisely because every token reads all the weights. So 'it runs 120B' is true, but read the architecture: MoE fast, dense slow.

Why does memory bandwidth matter more than the NPU's TOPS rating?

Because generating text is memory-bandwidth-bound, not arithmetic-bound. For each token the machine reads the entire set of active weights out of memory and does a relatively small amount of math on them - so bandwidth (GB/s) sets the speed, and unified-memory capacity sets which models fit at all. The NPU 'TOPS' figure measures raw arithmetic throughput, and mainstream local runtimes (llama.cpp, Ollama, MLX) run on the GPU, not the NPU. It is close to irrelevant for large-model text generation, which is why we lead with bandwidth.

Which platform has the best software support for running LLMs?

NVIDIA, then Apple, then AMD - today. CUDA and the DGX stack are what most AI tooling targets first, so the GB10 boxes have the smoothest path. Apple's MLX has matured fast and is now an excellent, well-optimised route on Apple silicon (plus Metal and llama.cpp). AMD's ROCm and Vulkan paths work with llama.cpp and are improving steadily, but they are still the roughest of the three and occasionally trail on day-one model support. If a painless software experience matters most, weight NVIDIA or Apple.

Should I buy one of these or just use cloud AI?

Buy one only if you run inference heavily or need your data to stay on-device. If you keep a coding assistant busy all day, run batch jobs, or want an always-on private agent, owning a $2,000-5,000 box pays off and keeps your prompts fully private. If your use is occasional, renting a cloud GPU by the hour or paying per token is usually still cheaper and avoids the maintenance. The strongest reasons to buy are privacy and independence from subscription pricing; the pure cost case only closes at high, sustained utilisation.

Related buyer's guides

All reviews →

Buyer's guide / AI PCs

LLM VRAM calculator

Enter a model size, quantization and context length to estimate the VRAM or unified memory an LLM needs to run locally — and instantly see which 2026 machines can hold it.

Use the calculator →

Buyer's guide / AI PCs

How to run LLMs locally: the 2026 hardware guide

How much VRAM or unified memory you need to run 7B-120B models locally, the capacity-vs-bandwidth tradeoff, and the 2026 hardware that hits each point on the curve — GB10, Ryzen AI Max, Apple Silicon and RTX GPUs.

Read the guide →

Buyer's guide / AI PCs

Mac vs PC for running local LLMs (2026)

It's not Mac vs PC — it's capacity vs speed. A Mac's unified memory fits models a PC GPU can't hold; a PC GPU runs the models that fit far faster. A sourced head-to-head plus which to buy.

Read the comparison →

Buyer's guide / AI PCs

The best AI PC for running LLMs locally in 2026

Six machines for running large language models on your own hardware — from a $799 Mac mini on-ramp to an NVIDIA GB10 supercomputer — ranked by what actually matters: unified-memory capacity and bandwidth, not NPU TOPS.

6 picks →

Buyer's guide / AI PCs

Best CPU for AI and local LLMs (2026)

The honest version: for pure GPU inference the CPU barely matters — so this guide is organised by when it does. Consumer value, the Intel option, unified-memory APUs that hold models a GPU can't, and the workstation platform you only need for multi-GPU rigs.

6 picks →

Benchmark / AI PCs

AI PC benchmark: local-LLM speed and memory (2026)

How the 2026 AI PCs compare for running local LLMs: unified memory, memory bandwidth, NPU TOPS, the biggest model each can hold, and measured tokens per second.

Read the benchmark →