Hardware compared

What to run local models on in 2026

The honest short version, and it is not the one you would expect from a company that rents Macs. If your model fits in 32 GB, buy an NVIDIA card. An RTX 5090 moves 1,792 GB/s against 546 for an M4 Max, and nothing Apple sells below the Ultra tier comes close. Above 32 GB the question changes completely, because speed stops mattering when the model will not load at all, and that is where unified memory earns its price.

Disclosure. We rent dedicated Apple Silicon, so we have an obvious interest in one of these three answers. That is exactly why the numbers below are computed the same way for all of them, why every specification links to its source, and why the section on where a Mac is the wrong purchase is the longest one on the page.

The market has split into three shapes

Fast and small. Discrete GPUs put GDDR7 behind a wide bus and reach between 936 and 1,792 GB/s. The catch is that consumer cards stop at 24 or 32 GB, so the models they can hold are limited to roughly 32B at Q4. Two cards double the capacity on paper but split the model across a bus, which is a different and worse thing than one pool.

Big and slow. The 128 GB unified memory boxes, the NVIDIA DGX Spark (GB10) and the AMD Strix Halo systems, use LPDDR5x. They will hold a 123B model, and they will do it at 273 and 256 GB/s respectively, which is between a quarter and a seventh of a modern GPU.

Big and fast, at a price. Apple’s Ultra tier is the only consumer hardware that is both. The M5 Ultra addresses 512 GB at 1,200GB/s, which is more memory than any of these and more bandwidth than all but the very fastest cards. NVIDIA’s answer in that quadrant is the RTX PRO 6000 at 96 GB and 1,597 GB/s, which is a workstation part with a workstation price.

The comparison that surprises people

The NVIDIA DGX Spark (GB10) costs $3,999 and has 273 GB/s of memory bandwidth. The M4 Pro, a two year old mid tier chip in a Mac mini, has 273 GB/s. They are the same number. The M4 Max (40-core GPU) has 546 GB/s at the same 128 GB of capacity, which is twice the Spark, and the M5 Ultra is 4.4 times it. If you are buying a 128 GB box for generation speed, that is the comparison to make before you spend the money. If you are buying it for CUDA, prompt processing, or fine tuning, none of the above applies and the Spark is a reasonable purchase.

Everything, at Q4_K_M

Tokens per second for generation, computed identically for every row from memory bandwidth and model size. A dash means the model does not fit in that hardware at all. Memory shown is the largest configuration available.

HardwareMemoryBandwidth8B32B70B123BPrice
GeForce RTX 3090Discrete GPU24 GB936 GB/s105~$1,300
GeForce RTX 409024 GB1,008 GB/s109varies
GeForce RTX 509032 GB1,792 GB/s14662varies
RTX PRO 6000 Blackwell96 GB1,597 GB/s1395730.118.2varies
NVIDIA DGX Spark (GB10)Unified memory box128 GB273 GB/s42.9125.73.3~$3,999
AMD Ryzen AI Max+ 395 (Strix Halo)128 GB256 GB/s40.711.35.43.1~$2,000
M5 UltraApple Silicon512 GB1,200 GB/s12045.523.313.9varies
M5 Max (614 GB/s)128 GB614 GB/s8025.512.57.3varies
M5 Pro64 GB307 GB/s47.313.46.4varies
M632 GB153 GB/s266.9varies
M4 Max (40-core GPU)128 GB546 GB/s7422.911.26.5varies
M4 Pro64 GB273 GB/s42.9125.7varies
M3 Ultra512 GB819 GB/s9732.916.49.7varies

Before you spend three thousand dollars on a guess

Everything above assumes you already know which model you need. Most people do not, and that is the expensive part: the row you are looking at costs thousands, and you cannot return it once you find out that a 70B at 5 tokens/sec is too slow for the thing you wanted it for, or that a 32B would have been fine all along.

A month on a rented M4 Pro with 64 GB costs $286. That is enough time to run your actual workload, on your actual documents, at the speeds in this table, and find out which row you should be buying. Then buy it.

And the case against us, since it is the same page. If you already know what you need, the machine sits on your desk, and you will still be using it in three years, owning wins on hardware cost and it is not close. We work through that arithmetic in full on the buy versus rent page, including the parts people forget: a static IP, power, a residential uplink, and being your own on call at three in the morning. Renting wins on uncertainty and on access, not on price.

Where a Mac is the wrong purchase

Where a Mac is the right purchase

Where the newest Apple chips land

Apple moved the whole line up a tier in 2026, and it changed the comparison more than the reviews suggested. Figures verified 2026-09-11.

ChipBandwidthMax memoryVersus the DGX Spark
M5 Ultra1,200 GB/s512 GB4.4x bandwidth
M5 Max (614 GB/s)614 GB/s128 GB2.25x bandwidth
M5 Pro307 GB/s64 GB1.12x bandwidth
M6153 GB/s32 GB0.56x bandwidth

Read the M6 row before drawing the wrong conclusion from the others. Apple’s entry desktop chip sits at 153 GB/s, which is below both 128 GB boxes, so the boxes are not simply slower than Macs. The crossover is the M5 Pro, the first chip in a Mac mini to edge past them at all, and it does so by twelve percent. What separates the classes is not the entry point but the top: Apple keeps scaling bandwidth to 614 and 1,200 GB/s while the boxes stop at 273. Their argument was never speed. It is 128 GB for around $2,000, and against that specific number Apple has nothing to offer.

How these numbers were produced, and what is wrong with them

Every speed on this page comes from one formula applied to every row: seconds per token equals the weight size divided by bandwidth times 0.915, plus 3.9 ms of fixed overhead. Those two constants were fitted by least squares to six models we measured ourselves on a base M4, and they reproduce all six to within 4%. The method and the raw data are on our benchmarks page, and the calculator will run it for any combination you like.

The obvious objection, stated before someone else does. That overhead constant was measured on Apple hardware running Ollama. Applying it to CUDA and ROCm assumes a comparable per token cost, which we have not verified and cannot until we measure those stacks. It is not a rounding error either: without it the formula claims an RTX 5090 generates over 300 tokens/sec on an 8B model, which nobody observes. With it the figure is 146, which matches what people report. We think that is the better assumption, but it is an assumption, and the faster the hardware the more the answer depends on it.

Everything else here is a published specification rather than a measurement. Sources for each piece of hardware:

  • GeForce RTX 3090 24 GB at 936 GB/s (used, $1,200 to $1,400 as of late 2025). Source
  • GeForce RTX 4090 24 GB at 1,008 GB/s (used, varies widely). Source
  • GeForce RTX 5090 32 GB at 1,792 GB/s (retail, varies). Source
  • RTX PRO 6000 Blackwell 96 GB at 1,597 GB/s (workstation class, quoted rather than listed). Source
  • NVIDIA DGX Spark (GB10) 128 GB at 273 GB/s (list price at launch). Source
  • AMD Ryzen AI Max+ 395 (Strix Halo) 128 GB at 256 GB/s (128 GB systems from a little over $2,000). Source

Street prices move constantly and we do not try to track them. The memory and bandwidth columns are the durable part of this comparison; treat the price column as an order of magnitude.

Common questions

Is a Mac or an NVIDIA GPU better for running local LLMs?

It depends entirely on whether your model fits in the GPU's VRAM. Below about 24 to 32 GB a discrete NVIDIA card wins and it is not close: an RTX 5090 moves 1,792 GB/s against 546 for an M4 Max, so it generates roughly three times faster on the same model. Above that ceiling the comparison inverts, because a 70B model at Q4 simply will not load on a 32 GB card while a 64 GB Mac runs it. Capacity decides what is possible and bandwidth decides how fast, in that order.

Is the DGX Spark faster than a Mac for local inference?

For generation, no. The DGX Spark has 273 GB/s of memory bandwidth, which is exactly the same figure as an M4 Pro and half that of an M4 Max. Since generation speed is set by bandwidth divided by model size, a Mac Studio generates about twice as fast at the same model. Where the Spark wins is prompt processing, which is compute bound rather than bandwidth bound, and the CUDA software stack, which matters enormously if you intend to fine tune rather than only serve.

What is the cheapest way to run a 70B model locally?

A 70B model at Q4 needs roughly 50 GB of memory once the context and the operating system are accounted for. The cheapest hardware that holds it is a 128 GB Strix Halo system at a little over $2,000, which will generate around 5 tokens per second. A 64 GB Mac reaches a similar speed. No consumer GPU holds it at all without splitting the model across two cards.

Why do unified memory boxes have so much less bandwidth than GPUs?

They use LPDDR5x, which is system memory designed for power efficiency, while GPUs use GDDR7 or HBM designed for throughput. That is the whole trade. LPDDR5x lets a machine address 128 or 512 GB in one pool at moderate speed; GDDR7 gives enormous speed across a pool that stops at 32 GB on consumer parts. Apple's Ultra chips are the exception that proves the rule, reaching 1.2 TB/s from LPDDR by using an extremely wide bus, and they are priced like it.

When the answer is not a box on your desk

There is one case this whole comparison cannot settle, because it is not about hardware. If the machine has to be reachable by your team or your application, around the clock, from anywhere, then buying the fastest box is the easy half and the hosting is the rest of it: a static IP, an uplink that does not drop, disk encryption, and someone awake when it hangs. We rent dedicated Apple Silicon from $99/mo with that part already done, Ollama installed and an OpenAI compatible endpoint exposed. What we own today is M4 generation; the M5 and M6 machines are listed in our configurator at real prices and route to a waitlist rather than to checkout, which is how we decide what to buy next.