M5 Ultra’s 512GB: Can a Desktop Hold a Frontier Model?

Apple’s M5 Ultra Mac Studio, announced August 25, 2026, offers up to 512GB of unified memory at 1.2TB/s — enough addressable memory to hold a frontier-class open-weight model for local inference on a desktop. That sentence has been the dream of the local-AI crowd for two years, and today it is roughly true. The interesting parts are the word “roughly,” the price Apple has not published, and the model that still does not fit.

The capacity threshold is real. DeepSeek-V4-Pro — 1.6 trillion total parameters — fits inside even the 256GB configuration you can order today. But Kimi K3, the single largest open-weight model in existence as of this writing, needs roughly 594GB even at 1-bit quantization. The headline claim fails on the actual largest model, and Apple's own flagship 512GB configuration is not orderable yet and carries no published price.

This post does one narrow job: it checks Apple's brand-new memory ceilings against the real disk footprints of four major open-weight models released before August 25, 2026, then runs the capex-vs-API break-even on published rates only. Capacity, bandwidth, and price — nothing else, because the rest of the local-AI buying decision is already covered elsewhere on this blog.

Apple’s newsroom release puts two chips in the new Mac Studio. The M5 Max carries an 18-core CPU, up to a 40-core GPU, and up to 128GB of unified memory at 614GB/s, from $2,499. The M5 Ultra is the headline: Apple’s first quad-die design, fusing two M5 Max processors over a next-generation UltraFusion interconnect into up to a 36-core CPU, up to an 80-core GPU with Neural Accelerators, a 32-core Neural Engine, and up to 512GB of unified memory at 1.2TB/s — which Apple describes as 50% more bandwidth than the previous generation. Pre-orders opened today in 30 countries; general availability begins September 22.

On AI, Apple’s silicon-focused companion release claims the M5 Ultra lets users “run huge LLMs with hundreds of billions of parameters entirely on device.” Treat that as what it is: vendor marketing copy. Apple attached no model name, no parameter count, and no tokens-per-second figure to the claim anywhere in either release. Whether the claim survives contact with real open-weight models is exactly what the rest of this post checks.

The same event refreshed the Mac mini around the 2-nanometer M6 from $899 — a capable desktop, not an inference machine, with memory bandwidth of 170GB/s on the 24GB and 32GB configurations and 153GB/s on the 16GB base. For how Apple's studio-class silicon stacked up against dedicated AI hardware before today, see our DGX Spark vs M5 Max vs RTX 6000 Blackwell comparison.

One number that does not exist anywhere in this post: tokens per second. No independent throughput measurement of the M5 Max or M5 Ultra existed on August 25, because the hardware had not shipped. Several aggregator sites already circulate specific figures attributed vaguely to “prior-generation” hardware or unlabeled community runs, with no reproducible method disclosed. When real benchmarks land after September 22, they will settle the throughput question — until then, anyone quoting a tokens-per-second number for these chips is guessing.

Here is the detail most launch-day coverage glossed over: the 512GB configuration — the one that makes the frontier-capacity argument possible at all — cannot be ordered today, and Apple has not said what it will cost. MacRumors confirmed that no price has been published for the tier. What you can order today tops out at 256GB of unified memory with 16TB of storage for $18,299.

That single unpriced line matters more than it looks. A capex-vs-API break-even is only as good as its capex number, and the exact configuration this launch’s headline argument depends on does not have one. Every piece of hardware math in this post therefore uses the two prices Apple has actually published — the $5,499 base and the $18,299 maximum orderable build — and treats the 512GB tier as what it is today: an announced capacity with an announced month and an unannounced cost. When Apple publishes the number in late October, the arithmetic below extends in five minutes. Until then, nobody can honestly tell you what frontier local inference at 512GB costs, because Apple has not said.

The test is simple: take four major open-weight models released on or before August 25, 2026, look up the smallest practical quantization of each in Unsloth’s published quantization tables, and check them against Apple's three new memory ceilings. A model “fits” only if the weights load with enough headroom left for the KV cache, the OS, and a usable context window — for the quantization and KV-cache math behind these footprint numbers, see our standing reference.

Read down the right-hand column and the announcement’s real shape appears. The 512GB tier holds three of the four models checked here, and that is why this machine matters. The 256GB configuration you can actually order today already holds DeepSeek-V4-Pro, a 1.6-trillion-parameter model, with roughly 94GB to spare. What no Apple memory tier holds is the last row.

Kimi K3 is Moonshot AI’s 2.8-trillion-parameter sparse mixture-of-experts model — 896 experts, 16 active per token, 104 billion active parameters, native vision, a 1M-token context — and since its weights landed on July 27 it has been the single largest open-weight model in existence. Per Unsloth’s quantization tables, its smallest practical GGUF — a ~1.66-to-1.92-bit dynamic mix — is 594GB on disk and wants roughly 610GB of combined memory to run. Even shredded down to roughly one bit per weight, the model is 82GB too large for the biggest Mac Apple will sell this year.

This is where Apple’s “hundreds of billions of parameters” line deserves precision. Kimi K3’s active parameters — 104B — sit comfortably inside that phrase. But mixture-of-experts inference requires the full expert set resident in memory: the 2.8T-parameter total footprint is what has to fit, not the 104B slice that fires per token. By active-parameter count the vendor claim reads fine; by what actually determines whether a model loads, the largest open model overshoots the largest Mac by a comfortable margin. For a sharp independent read on what K3’s scale means for open weights generally, Interconnects’ analysis is the reference.

Capacity decides whether a model runs at all; memory bandwidth largely decides how fast it generates once it does, because autoregressive decoding re-reads the active weights for every token. And on bandwidth, the datacenter comparison is not close. NVIDIA’s H200 moves 4.8TB/s through 141GB of HBM3e — four times the M5 Ultra’s 1.2TB/s, and nearly eight times the M5 Max’s 614GB/s. One caveat worth carrying: NVIDIA’s own product page still labels those H200 numbers “preliminary specifications.”

The two-way cut, stated plainly: the Mac wins on capacity — an H200 carries 141GB against the Mac’s 512GB ceiling — and datacenter silicon wins on bandwidth, which is what turns into tokens per second. That trade cannot be put on a per-dollar footing today: NVIDIA publishes no street price for the H200 (only OEM and cloud partners quote it, and it varies widely), and Apple has published no price for the 512GB tier. A model that merely fits is not the same as a model that runs at production speed.

How that trade-off resolves for the new chips is, for now, unknowable — the hardware ships September 22, and no independent measurement exists. What history suggests is direction, not magnitude: unified memory buys you models that datacenter cards need multi-GPU rigs to load, and costs you generation speed against those same rigs. Where the crossover sits for the M5 generation is precisely the benchmark to wait for.

Here is the arithmetic, built strictly from numbers that were published on August 25. The hardware side is the $18,299 maximum orderable Mac Studio — M5 Ultra, 256GB unified memory, 16TB SSD — because it is the priciest configuration with an actual price, and because it genuinely holds DeepSeek-V4-Pro, the largest open-weight model that both fits it and has a fully published metered rate. The API side is DeepSeek’s own price list: V4-Pro at $0.66 per million input tokens and $1.98 per million output tokens, off-peak, cache-miss.

Thirteen-point-nine billion tokens is a lot — and whether it is a lot for you is the entire question. A team running continuous multi-hour agent workloads can clear that volume in months; a team making occasional API calls never will. The deeper versions of this decision are already worked through in our local-vs-cloud subscription ROI analysis, the running-cost side in local workstation economics, and the fuller self-hosting picture in our frontier self-hosting TCO analysis — this post’s contribution is the new capacity threshold those analyses could not price, because it did not exist.

Map the model you intend to run to the cheapest memory tier that holds it with headroom, and the buying decision mostly makes itself. Four honest lanes, from the quantization table above:

Two forward-looking notes. First, the software side is not the bottleneck it was: Apple’s MLX stack is what this hardware is built to run, and our MLX framework guide covers the toolchain that turns unified memory into a working inference box. Second, expect the capacity race to keep moving under this machine: the largest open-weight total nearly tripled in roughly a year — Kimi K2 shipped at 1T parameters in 2025, K3 at 2.8T this July — and there is no reason to assume K3 stays the ceiling. A 512GB Mac that holds most of the models checked here may hold fewer of next year’s — which argues for buying against the model you will actually run, not against the frontier as a category. If you are weighing local inference against metered APIs for production workloads, our AI transformation engagements start with exactly this fit-and-economics evaluation on your own workload numbers.