All three boxes put about 128GB of memory behind one GPU. That’s why they share a shortlist, and why you’ll see DGX Spark vs Strix Halo asked in both directions.
NVIDIA moved the price on October 2, lifting the 128GB DGX Spark to $6,950, and that changes the answer.
Below you will find the price per GB and the usable memory, cited tokens per second on gpt-oss-120b and a dense 70B, and the software each one locks you into. Amazon prices and stock were checked on October 19, 2026, and vendor list prices in early October 2026.
Which one to buy
For most people running local models, buy a 128GB Strix Halo box. It costs about half a DGX Spark, and on gpt-oss-120b the Spark decodes only 8 to 20 percent faster.
Best value: Bosgame M5, Ryzen AI Max+ 395, 128GB. It’s the cheapest 128GB Strix Halo on Amazon at around $3,600, though stock runs low, so check the live price.
Best for CUDA work: NVIDIA DGX Spark, 128GB. Prompt processing runs roughly five times faster than Strix Halo. NVIDIA raised its list price to $6,950 on October 2, and its DGX Spark page sends buyers to NVIDIA Marketplace, where NVIDIA’s own 128GB unit has been reported out of stock and partner GB10 systems run about $7,000 to $9,000, so check stock before you plan around it. The ASUS Ascent GX10 is built on the same GB10 chip with the same 128GB and ConnectX-7 networking, but every GX10 listing on Amazon was out of stock in mid-October.
Fastest for one user: Mac Studio, M5 Max with the 40-core GPU and 128GB. It has 614 GB/s of bandwidth, so it generates tokens fastest, and Apple sells this configuration at about $5,100.
Amazon does not stock the 128GB Mac Studio. It does sell the base M5 Max Mac Studio with 36GB for about $2,300, which is enough for models up to about 27B.
How we compared them
None of these three has been on our bench, so every speed below is taken from a published run, linked at the figure.
Specs were taken from the NVIDIA, AMD and Apple pages. Amazon prices and stock were read from Amazon’s catalog API on October 19, and NVIDIA and Apple list prices between October 2 and October 13. Runtimes differ between rows, so read each comparison inside its row, not across the whole table.
Two numbers are our estimates, and they are marked as such. Memory bandwidth gets its own column because, for one model at a time, it sets the ceiling on token speed.
Strix Halo vs DGX Spark vs Mac Studio prices
Strix Halo is the cheapest memory you can buy behind a GPU right now, and the Mac gives you the most bandwidth per dollar.
| Machine | Price (early Oct 2026) | Memory | Bandwidth | $ per GB | GB/s per $1,000 |
|---|---|---|---|---|---|
| Strix Halo, Ryzen AI Max+ 395 (Bosgame M5) | about $3,600 on Amazon | 128GB | 256 GB/s | $28 | 71 |
| DGX Spark 128GB (NVIDIA) | $6,950 | 128GB | 273 GB/s | $54 | 39 |
| DGX Spark 64GB (partners, from Oct 23) | $4,999 | 64GB | 273 GB/s | $78 | 55 |
| Mac Studio M5 Max, 40-core GPU | $5,099 | 128GB | 614 GB/s | $40 | 120 |
| Mac Studio M5 Ultra (base) | $5,499 | 96GB | 1.2 TB/s | $57 | 218 |
The base M5 Ultra has only 96GB, so it holds less than the 128GB boxes despite its bandwidth. The 256GB and 512GB options need the 80-core GPU chip and cost well above the base price. Other 128GB Strix Halo boxes cost a little more: the GMKtec EVO-X2 is about $3,650 and stays in stock. Our Ryzen AI Max+ 395 mini PC comparison covers the rest by ports and noise.
A 128GB DGX Spark now costs $3,351 more than the Bosgame. It launched at $3,999, and memory prices drove the rise, which The Register puts at nearly 75 percent in a year.
Tokens per second
On one chat stream, the Mac generates fastest, the Spark and Strix Halo sit close together, and the Spark wins prompt processing by a wide margin.
| Workload | Strix Halo 395 | DGX Spark | M5 Max, 40-core |
|---|---|---|---|
| gpt-oss-120b, generate | 54.5 (Vulkan), 49.0 (ROCm) | 58.7 | 87.9 (MLX, 4K context) |
| gpt-oss-120b, prompt processing | 459 (512 tokens) | 2,444 (2,048 tokens) | 1,325 (4K context) |
| Dense 70B class, Q4/Q5, generate | 4.1 (Qwen2.5 72B Q5_K_M) | 4.4 (Llama 3.1 70B Q4_K_M) | about 11 (estimate) |
| 7B to 8B, Q4, generate | 43.6 (Llama 2 7B) | 38.0 (Llama 3.1 8B, Ollama) | 119.9 (Llama 2 7B Q4_0) |
Sources: Strix Halo from a Vulkan run and a ROCm run on an EVO-X2. Spark from the llama.cpp DGX Spark bench page and Ollama’s Spark numbers. M5 Max from Hardware Corner, measured on a MacBook Pro with the same chip, and the llama.cpp Apple Silicon table.
Generation speed follows bandwidth, so the Mac, with 2.25 times the Spark’s bandwidth, generates gpt-oss-120b 1.5 times faster.
Prompt processing follows compute instead, and that is where the GB10 earns its price. It ran about 5.3 times faster than Strix Halo on gpt-oss-120b, though the two runs used different prompt lengths.
Long prompts change the Mac picture, though. At 16K context the M5 Max processed 2,710 tokens per second in MLX while the Spark managed 1,956 in llama.cpp, and because the runtimes differ, read it as a trend.
The 70B estimate comes from the M5 Max’s measured 7B speed, scaled by bandwidth to a 42.5GB Q4 file. Treat it as a ballpark until someone posts a run.
Serving several users
Batched serving is the Spark’s strongest case. In StorageReview’s prefill-heavy vLLM run on gpt-oss-120b (8K-token prompts, 64 concurrent requests), the Spark pushed 2,760 tokens per second against 314 for AMD’s own Ryzen AI Halo box.
That is 8.8 times faster, and on this model long prompts are where the gap is widest. With equal-length prompts on the same model it shrank to 3.2x, and the reviewer puts the usual vLLM gap at 2x to 4x. If you plan to share one box across a team, or run several agents against it, that ratio matters more than single-stream speed.
TensorRT-LLM only runs on NVIDIA hardware, so of these three only the Spark can use it. How much the engine choice matters shows in our TensorRT-LLM vs vLLM benchmark on an RTX 5090.
How much memory models get
None of these gives a model the full 128GB, and each one caps it differently.
- DGX Spark: about 120GB is usable for models, according to Ollama. No tuning needed.
- Strix Halo: Windows tops out at 96GB set in the BIOS. On Linux, a kernel parameter raises the GPU’s share to about 115 to 120GB.
- Mac Studio: macOS lets the GPU wire about 75 percent of memory, roughly 96GB of 128GB. A sysctl raises it, but it resets on every reboot.
The kernel parameter and the sysctl are covered in our guide to the best mini PC for local AI, under opening the full memory pool. For sizing a model against those limits, the LLM VRAM calculator does the arithmetic.
All three run gpt-oss-120b, which needs about 63GB, but none of them runs a dense 70B fast. If you need more than 120GB, the 256GB M5 Ultra is the only single box here that goes past it.
DGX Spark: CUDA on Arm
Buy the Spark for the software, not the token speed.
DGX OS ships with CUDA and a container runtime, and PyTorch, vLLM and TensorRT-LLM run from NVIDIA’s NGC containers. Blackwell also gives you NVFP4, which made Qwen3 8B 30 percent faster than FP8 on an RTX 5090 in our TensorRT-LLM tests, at a 1.5-point accuracy cost. Fine-tuning code written for an NVIDIA GPU moves over with few changes.
Its CPU is a 20-core Arm chip. Any container image or Python wheel built only for x86 will not run, so check your stack before you order. The two QSFP ports carry 200 Gbps ConnectX-7, and up to four units can be clustered, though more than two need a QSFP switch.
Not worth it for one person chatting with one model. You’d pay about $3,350 extra for roughly 8 percent more decode speed than the Bosgame. The same GB10 chip also ships in laptops, covered in our RTX Spark explainer.
Strix Halo: x86 and cheap
Strix Halo is a normal x86 PC that happens to have 128GB of fast shared memory.
Windows, Ubuntu, Fedora and Proxmox all install on it. llama.cpp works over Vulkan or ROCm, and Vulkan came out ahead in the two runs above. Under load the boxes draw roughly 144 to 186 W, depending on the maker’s power limits.
Prompt processing is its weak spot. At 459 tokens per second, a 20,000-token codebase prompt takes over 40 seconds before the first word comes back, where the Spark’s long-context rate puts the same prompt near 10 seconds.
Look elsewhere for long prompts, RAG over big documents or batched serving. Networking is also thin: the Bosgame has a single 2.5GbE port. If you want 192GB, AMD’s newer chip costs roughly double, as our Ryzen AI Max+ 495 vs 395 comparison shows.
Mac Studio: bandwidth first
The Mac Studio is the quickest of the three at turning out tokens.
Pick the 40-core GPU. The 32-core M5 Max has 460 GB/s and tops out at 36GB, so it’s a different machine for this job. With 128GB the 40-core version is $1,851 cheaper than a 128GB Spark and generates 1.5 times faster.
You run MLX or llama.cpp on Metal. CUDA code does not come across, and fine-tuning tools are thinner than on NVIDIA. Memory is fixed at purchase, and the 10GbE port is the fastest wired network unless you cluster over Thunderbolt 5.
Wrong pick when your work is CUDA, or when you’d rather run Linux. For smaller agent models, a Mac mini is the cheaper Apple route, compared in our Mac mini for AI agents guide.
Pick by workload
Match the box to what you’ll run most of the time:
- One user chatting with MoE models like gpt-oss-120b on a budget: Strix Halo.
- One user who wants the fastest replies and doesn’t need CUDA: Mac Studio M5 Max, 128GB.
- CUDA development, fine-tuning, long prompts or a shared inference server: DGX Spark or the ASUS GX10.
- Models larger than 120GB in one box: Mac Studio M5 Ultra with 256GB.
If none of these fits your budget, a discrete card is still the faster route for models that fit in its VRAM. Our best GPU for LLMs guide covers those.
What changes after October 23
Two launches this month shift the comparison, so check them before you order.
Sales of the 64GB DGX Spark start on October 23, from $4,999, through Acer, ASUS, Dell, Gigabyte, HP and MSI. Bandwidth is kept at 273 GB/s, so token speed on models that fit should match the 128GB unit.
NVIDIA rates it for models up to 100B, while The Register frames its sweet spot as 26B to 35B. Two of them cluster to 128GB, though at $9,998 that costs more than one 128GB Spark.
Apple lists a 512GB M5 Ultra as coming late October, with no price yet. That doubles the largest model you can hold in one box.
Strix Halo prices have also moved up through the year. If the Bosgame shows out of stock, the EVO-X2 is the next cheapest 128GB box and runs the same chip. For month-to-month moves on cards and boxes, our GPU prices tracker is updated each month.
