Mac Studio M5 Ultra: 512GB at 1.2TB/s, what it changes for a local LLM

Apple announced the Mac Studio M5 Ultra on August 25, 2026. The machine moves to 1.2TB/s of unified memory bandwidth, keeps a 512GB ceiling and starts at $5,499. Here is what those numbers actually buy : and what nobody can claim yet, for lack of any measurement.

By QDNA · Published August 28, 2026 · 11 min read

What Apple announced, official figures

Apple offers two chips with distinct positions: the M5 Max targets the creative workstation, whereas the M5 Ultra is the one that matters for an inference platform, since it remains the only variant that goes beyond 128GB of unified memory.

Sources: Apple Mac Studio technical specifications and the Apple Newsroom announcement of August 25, 2026. European pricing starts at €2,999 and €6,599 incl. VAT respectively, recorded on the French Apple Store on August 28, 2026.

Two details in the configuration grid matter more than all the rest, and both are easy to miss when reading a product sheet.

The 256GB and 512GB options are not available on the entry-level configuration. Apple reserves them for the variant fitted with the 36-core CPU and the 80-core GPU, so that a machine billed at $5,499 cannot be taken to 512GB at all: you first have to move up to the higher chassis before paying anything for memory.

Bandwidth does not depend on how much memory you order. The figure of 1.2TB/s is quoted for both M5 Ultra variants, including the 96GB one. That is the opposite of the M5 Max, whose bandwidth only rises from 460 to 614GB/s with the 40-core GPU. For LLM work it means a 96GB M5 Ultra delivers the same generation throughput as a 512GB one : on the models that fit in 96GB.

1.2TB/s: the only figure that decides output throughput

Serving a language model means two distinct regimes, and they do not depend on the same component. We covered this in prefill, decode and KV cache: which hardware for an LLM; the M5 illustrates the split almost caricaturally.

Apple writes it plainly in its own research publication: "generating the first token is compute-bound", while "generating subsequent tokens is bounded by memory bandwidth, rather than by compute ability". The Neural Accelerators built into the M5 GPU therefore massively speed up the prefill phase, and much less the generation phase.

Apple's published measurements, M5 against M4, give the order of magnitude:

Source: Apple Machine Learning Research, "Exploring LLMs with MLX and the Neural Accelerators in the M5 GPU". Method: 4,096-token prompt, 128 tokens generated, MLX. These measurements compare an M5 to an M4 on MacBook Pro, not an M5 Ultra to an M3 Ultra. They describe the behaviour of the architecture, not the performance of the machine this article is about.

The ratio is spectacular in one column and modest in the other: up to four times faster on the first token, a fifth to a quarter better on everything after. And the gap is explained entirely by bandwidth: between M4 and M5 it rises from 120 to 153GB/s, or +28% : precisely the order of magnitude of the observed generation gain.

That is precisely what makes the figure announced for the M5 Ultra interesting, since going from 819GB/s to 1.2TB/s represents 47% more of the quantity that bounds generation. On a machine whose memory capacity does not move at all, that is where the whole of the progress sits.

The generation ceiling, calculated and not measured

Nobody has measured an M5 Ultra yet, since the machine only ships from September 22, 2026 and the 512GB configuration will not arrive before late October. The tokens-per-second figures already circulating are therefore estimates, sometimes presented as readings, which is why we do not reproduce them here.

What can be calculated, on the other hand, is the ceiling that the hardware cannot exceed. Generating a token means reading the weights that take part in the computation: all of them for a dense model, only the active experts for a mixture-of-experts model. Dividing bandwidth by that volume yields an upper bound the hardware cannot exceed : in practice 50 to 80% of it is reached, depending on runtime, context and concurrency.

QDNA calculation: announced bandwidth ÷ volume of weights read per token. These are arithmetic ceilings, not measurements, they include neither attention-cache reads, nor activations, nor runtime overhead. The table only lists models that actually fit on the machine and whose active parameter count is published: Qwen3.8-Flash-Next (125B, 6B active), DeepSeek V4 Flash 0731 (284B, 13B active), GLM 5.3 Flash (320B, 18B active) and Qwen3.8-27B (dense). Very large mixture-of-experts models such as Kimi K3 are excluded: they fit in no format, so no throughput figure means anything for them here.

On the same arithmetic, an M3 Ultra at 819GB/s would top out around 273 tokens/s for Qwen3.8-Flash-Next in Q4, against 400 for the M5 Ultra; 126 against 185 for DeepSeek V4 Flash 0731. The ratio of 1.47 recurs line by line, since it depends on nothing but bandwidth.

Which models fit, depending on the memory ordered

Capacity, by contrast, has not moved since the M3 Ultra and 512GB remains the ceiling, so that the question of which model fits has the same answer as before and depends entirely on the memory option chosen at order time.

QDNA calculation: parameters × bytes per parameter × 1.15 runtime margin, excluding attention cache. Same rule as the site's sizing matrix. Sizes and active parameter counts taken from the official model cards: Qwen3.8-Flash-Next (125B, 6B active, 262,144 native context extensible to 1M), DeepSeek V4 Flash 0731 (284B, 13B active) and GLM 5.3 Flash (320B, 18B active, MIT licence). Accessed August 28, 2026.

The verdict has reversed with the current generation of open models, for a reason that lies in their architecture. "Flash" mixture-of-experts architectures activate few parameters per token, which makes them both compact once quantised and fast: the 96GB entry configuration already runs Qwen3.8-Flash-Next in Q4, not just a dense 27B model. It is 256 and 512GB that unlock DeepSeek V4 Flash and GLM 5.3 Flash, and they alone justify moving to the 36-core variant.

What stays out of reach are models with a very large total parameter count, whatever their active count: Kimi K3 (2,800B) needs roughly 1,610 GB of weights in Q4. That is the machine's structural limit : 512GB is a lot for a workstation, and little against a DGX Station or an RTX PRO 6000 server.

MLX: what 512GB actually buys

On Apple Silicon the choice of inference runtime directly conditions the performance observed. The comparative study published by Rajesh et al., conducted on a 192 GB M2 Ultra Mac Studio with the Qwen-2.5 family and prompts ranging from a few hundred to 100,000 tokens, concludes that MLX achieves the highest sustained generation throughput, while MLC-LLM delivers lower time-to-first-token on moderate prompts, and llama.cpp stays highly efficient for lightweight single-stream use. The authors also state the limit honestly: these stacks "still trail NVIDIA GPU-based systems such as vLLM in absolute performance".

That hierarchy matches what Apple measures on its own side, since the Neural Accelerators speed up the prefill phase while generation remains bounded by memory bandwidth, and MLX is the runtime that exploits both dimensions together.

The quantisation ladder, measured

This is where memory capacity stops being a spec-sheet line and becomes a quality decision. The MLX conversions of GLM 5.3 Flash published by orcarouter come with perplexity and token-agreement measurements against the FP8 reference:

Source: orcarouter/GLM-5.3-Flash-MLX, perplexity measured against the FP8 reference at 2.7797. The "minimum machine" column is our reading: size on disk plus headroom for the attention cache and runtime state.

The reading of this table is as blunt as it is instructive for anyone about to place an order. Between 6 and 4 bits, quality loses just under three points of perplexity while token agreement stays above 96%, which makes the gap real but arguable. Between 4 and 2 bits however, perplexity rises by 57% and agreement falls to 86.6%. The author of these conversions says it plainly for the most aggressive variant: "long code generation is not reliable at this precision", with repetition loops and missing glue code.

In other words, 512GB does not buy you the ability to run one more model, it buys you the right to run it at 6-bit rather than 4-bit. On writing or question-answering that gap will not show at all, whereas on long code generation it sits exactly where the line falls between a usable assistant and one that invents.

GLM 5.3 Flash on M5 Ultra, the concrete case

Crossing the three preceding sets of figures produces the table a buyer actually wants, since the MLX variant chosen determines at once the machine required, the throughput reachable and the quality retained. GLM 5.3 Flash activates 18 billion parameters out of 320, which makes its generation ceiling a direct function of the precision selected: fewer bits per parameter means fewer bytes re-read for every token, therefore more tokens per second, at the price of a degradation that perplexity measurements quantify precisely.

Sizes, perplexity and token agreement taken from orcarouter/GLM-5.3-Flash-MLX. The ceilings are a QDNA calculation, announced bandwidth divided by the 18 billion active parameters at the precision considered, and not a measurement. The "minimum machine" column adds headroom for the attention cache and runtime state to the model size.

The practical reading fits in one sentence: the 512GB configuration is the only one that allows GLM 5.3 Flash to be served at 6-bit, meaning at a quality practically indistinguishable from the FP8 reference, whereas 256GB forces a drop to 4-bit and three additional points of perplexity. The move is counter-intuitive for anyone reasoning in throughput, since dropping precision accelerates generation: 4-bit tops out at 133 tokens per second against 89 at 6-bit. In other words, 512GB does not buy speed, it buys the right to decline speed bought at the expense of quality.

One practical constraint remains, which the product sheet does not mention: at the moment the machine is delivered, serving this model requires installing mlx-vlm from its main branch, where the glm5_next architecture only landed on August 26, 2026, together with the third-party patch that restores numerical parity with the reference.

The real state of support

One point product pages do not mention and which decides short-term feasibility: the glm5_next architecture only landed in the mlx-vlm main branch on August 26, 2026, and no published release carries it yet. PipeNetwork's 4-bit conversion, at 177.6 GB against 642.7 GB for the upstream bfloat16, documents in passing two numerical bugs and two epsilon mismatches found in that branch, fixed by a patched runtime reaching 1e-6 end-to-end parity with the reference.

This is not an implementation detail, because it means that at the moment the machine ships, running the model that most interests it requires installing mlx-vlm from the main branch together with a third-party patch. The operation remains doable, but it is not what a commercial datasheet describes.

The four-machine cluster

Apple states that "a four Mac Studio cluster delivers up to 3x faster AI inference performance", relying on Thunderbolt 5 and RDMA. The M5 Ultra offers six Thunderbolt 5 ports at 120Gb/s, two of them on the front.

The figure deserves to be read for exactly what it states. Four machines for a factor of three is 75% efficiency, which is honest for pipeline parallelism over an external link : but 120Gb/s is 15GB/s, roughly 1.2% of internal memory bandwidth. The link between machines is therefore three orders of magnitude slower than local memory: the split has to be done by layers, with very little exchange, and the gain lands on aggregate throughput, not on single-request latency.

For comparison, the UltraFusion interconnect joining the four dies inside a single M5 Ultra exceeds 4.4TB/s. The gap between the inside and the outside of the enclosure is what makes the cluster useful for serving several requests in parallel, and of little use for speeding up one.

Against the alternatives

The Mac Studio has never won on raw throughput, and the M5 Ultra does not change that. Its ground remains memory per euro and per watt, in a silent 3.6 kg enclosure drawing 480W at most.

• Against a DGX Spark (128GB, unified LPDDR5X): the 512GB Mac Studio plays in another capacity class, but the Spark stays inside the CUDA ecosystem, which counts the moment you want to fine-tune a model.
• Against a DGX Station (748GB, HBM3e and LPDDR5X): the Station loads models the Mac cannot, and serves more simultaneous requests. It also costs an order of magnitude more.
• Against an API: the question is not throughput but control. A Mac Studio transmits nothing to anyone, which settles data location in one move : see on-premises or API: comparing the real cost.

For an inference workstation dedicated to a single team, installed in an office rather than a server room, the M5 Ultra is today the densest memory option that requires no dedicated infrastructure. That scope remains narrow, but it corresponds to a need that organisations actually have.

What we do not know yet

Three unknowns remain open, and it is better to name them than to fill them with estimates.

• Real throughput, no independent measurement exists before the first deliveries on September 22. The ceilings above say what the hardware cannot exceed, not what it reaches.
• The price of the 256GB and 512GB options, apple does not publish them on its specifications page, and the figures circulating in the trade press contradict each other by a factor of two. A quote is required.
• How MLX behaves on a four-die chip, apple's published measurements cover a monolithic M5. How MLX spreads the load across the 80 GPU cores of an M5 Ultra, and at what efficiency, remains to be observed.

We will complete this article with readings as soon as a machine can be measured, following the same method as the measurements section: one (model, format, hardware) triple per page, with a reproducible procedure, and no value for anything that has not been measured.

What to remember

The M5 Ultra is a bandwidth update rather than a capacity update, which determines who it is for. The additional 47% feeds generation throughput directly, since throughput depends on bandwidth linearly, while the question of which model fits in memory gets exactly the same answer as with an M3 Ultra. An organisation whose constraint was capacity will therefore see no improvement, whereas one limited by speed at constant capacity gets the clearest step Apple has offered on this line.

The commercial trap lies in the entry configuration, since at $5,499 the machine carries only 96GB and cannot later be taken to 512GB. Sizing is therefore decided before the order is placed rather than after the machine is delivered. We help set it on the model × hardware matrix, or in conversation.