Mac Studio M5 Ultra 512GB: reported at $22,299, with 1.2TB/s of bandwidth, and which of the biggest open models actually fit

The Mac Studio M5 Ultra with 512GB of unified memory went on sale on October 28, 2026, MacRumors reported, completing the lineup Apple announced on August 25 with the promise that the top memory option was "coming in late October." MacRumors put US pricing for the 512GB configuration at $22,299, and Cult of Mac reported the same total. That buys the largest pool of memory Apple sells in one machine and 1.2TB/s of bandwidth, enough to hold some of the biggest open-weight models on a desk. But not all of them: Apple's own tests show that a model of about 800 billion parameters at 4-bit already overflows a single 512GB machine. This explainer applies the memory math from our M6 MacBook Pro explainer and our guide to choosing a model for your laptop to the top of the range.

This explainer is based on Apple's August 25, 2026 newsroom post, the Mac Studio product and tech-specs pages, Apple Machine Learning Research's October 15, 2026 post on distributed inference with MLX, Apple's online store, MacRumors' and Cult of Mac's October 28, 2026 reports, and Hugging Face model cards and repositories, checked October 29, 2026. Apple's performance figures come from Apple's own testing and are labeled vendor-reported; the memory estimates are illustrative calculations, not measurements.

The short version

• What's new: the 512GB memory option for the M5 Ultra Mac Studio, on sale from October 28, 2026, according to MacRumors.
• Price: $22,299 in the US for a 512GB configuration, as MacRumors and Cult of Mac reported. Apple's newsroom post lists M5 Ultra models from $5,499, and Apple's store lists the 36-core CPU, 80-core GPU model that 512GB requires from $6,799.
• Bandwidth: 1.2TB/s, which Apple says is 50% higher than before.
• What fits on one machine: on paper, 4-bit weights for about 900 billion parameters (512 × 8 ÷ 4.5); in practice much less. Reflection Beam (501B) fits comfortably; DeepSeek V4.1 Flash 1017 (763B) is tight.
• What doesn't: Meta's 816B Muse Spark Open at 4-bit, according to Apple's tests, and Kimi K3.1 Pro (2.8 trillion parameters), which needs about four machines.

What is the Mac Studio M5 Ultra with 512GB?

It's the top memory configuration of the Mac Studio Apple announced on August 25, 2026. According to Apple's tech-specs page, the M5 Ultra starts with a 30-core CPU, a 64-core GPU and 96GB of memory, and the 256GB and 512GB options require the upgraded 36-core CPU and 80-core GPU.

Last verified: October 29, 2026

Apple didn't publish a press release for the 512GB option. On October 29, the footnote on Apple's Mac Studio page still read "512GB memory option for M5 Ultra coming late October." Cult of Mac broke the $22,299 down: the $6,799 model with the 36-core CPU and 80-core GPU, plus $4,000 to reach 256GB and another $11,500 for 512GB. Apple's store lists the $6,799 starting price; the memory upgrade prices are as reported.

The option has a history. MacRumors reported in March 2026 that Apple had quietly dropped the 512GB upgrade, then a $4,000 add-on, from the previous M3 Ultra Mac Studio, which it linked to the global DRAM shortage. That shortage has also pushed up PC memory prices, as our Q3 PC shipments story explains.

What does Apple claim for AI performance?

All of these are vendor-reported, from Apple's tests in July 2026 unless noted:

• Peak AI compute: "up to 4.3x" that of the M3 Ultra.
• Prompt processing in LM Studio: up to 4x faster than the M3 Ultra and 9.8x faster than the M1 Ultra, measured as time to first token with an 8K-token prompt and a 14-billion-parameter model at 4-bit.
• Clusters: four Mac Studios connected over Thunderbolt 5 with RDMA deliver "up to 3x faster" distributed inference than one, in Apple's test with a 72-billion-parameter dense model on a 32K-token summarization task.

As H24 explains, time-to-first-token figures measure how fast a Mac reads your prompt. How fast it writes the answer depends mainly on memory bandwidth.

How much memory do the biggest open models need?

Use M7's formula: memory for weights ≈ parameters × bits per weight ÷ 8, with 4.5 bits for common 4-bit formats, 8.5 for 8-bit and 16 for BF16. Mixture-of-experts models only use a fraction of their parameters per token, but every expert has to sit in memory. The results below are illustrative calculations, not measurements, and leave out the attention cache, the runtime and macOS.

Last verified: October 29, 2026

The calculations work out as follows:

• Reflection Beam: 501.3B × 4.5 ÷ 8 ≈ 282GB, close to the 294GB NVFP4 repository our Beam story describes. That leaves more than 200GB for the attention cache and the system. The 8-bit version, at 503GB of files, doesn't leave room to run. Reflection's 4-bit weights are in Nvidia's NVFP4 format, so check that your Mac runtime supports it or use a converted build.
• DeepSeek V4.1 Flash 1017: it uses only about 16B parameters per generated token, but all 763B must be loaded. 763B × 4.5 ÷ 8 ≈ 429GB, leaving roughly 80GB for everything else. DeepSeek's published weights take 510GB, as our V4.1 Flash 1017 story reports, which fills the machine on their own.
• Muse Spark Open: Apple measured about 493GB in NVFP4 and wrote: "Add the memory needed for intermediate computations during inference, and the model cannot fit on a single machine." That's the best evidence that the practical ceiling sits well below 512GB. Meta's release suggested two 512GB or four 256GB Mac Studios, as our Muse Spark Open story reports.
• Kimi K3.1 Pro: Artificial Analysis lists 2.8 trillion total and 104 billion active parameters. 2.8T × 4.5 ÷ 8 ≈ 1,575GB, more than three machines' worth of memory, so in practice four. Moonshot hasn't released the weights yet; it promised them by November 4, 2026, as our Kimi K3.1 Pro story reports.

M7's rule of thumb is to budget about two-thirds of memory for a model you run on a machine you also work on. Two-thirds of 512GB is about 341GB, so on a Mac Studio that doubles as a workstation, Beam at 4-bit is the only one of these four that fits within it.

Does context length change what fits?

Yes. Every token of context adds to the attention cache, which M7 shows can rival a model's weights at long lengths. Apple's October 15 post gives two measured examples on the M5 Ultra: Qwen 3.8 27B in BF16, about 54GB of weights by the formula, used up to 68GB at 65,536 tokens, and Mistral Medium 3.5 128B in FP8, about 128GB of weights at one byte per parameter, used up to 158GB. Apple said both fit "comfortably" on a single 512GB machine. For the 500B-plus models above, plan on tens of gigabytes more for long contexts; the exact amount depends on each model's attention design.

How fast will big models run on it?

Bandwidth sets a ceiling. Each new token requires reading the active weights from memory, so dividing 1.2TB/s by the bytes read per token gives an upper bound. For Beam at 4-bit, 23.4B active parameters × 4.5 ÷ 8 ≈ 13.2GB a token, which gives a ceiling of about 91 tokens a second. Real speeds are lower, and attention, the KV cache and software overhead all cut into it. This is arithmetic, not a measurement.

Apple's own numbers show the gap. In its October 15 tests, a single M5 Ultra generated 7.70 tokens a second with Mistral Medium 3.5 128B in FP8 at an 8,192-token context, against a ceiling of about 9.4 by the same arithmetic (1,200GB/s ÷ 128GB). Muse Spark Open ran at 24.99 tokens a second on two M5 Ultras and 36.73 on four, using tensor parallelism. These are vendor-measured results on Apple's chosen setups.

Is one 512GB Mac Studio better than a cluster?

It depends on the model. One 512GB machine holds a 4-bit Beam or a 4-bit DeepSeek V4.1 Flash without the complexity of networking. Models above about 800B parameters at 4-bit need a cluster, which Apple supports over Thunderbolt 5 with RDMA in macOS 26.2 and MLX. Apple's tests show clusters also speed up models that would fit on one machine: Qwen 3.8 27B went from 19.69 tokens a second on one M5 Ultra to 50.89 on four.

For trillion-parameter models such as Kimi K3.1, the alternatives are data-center hardware. Cerebras, for example, describes spreading Kimi K3 across 59 wafers, as our Cerebras story explains, or software that streams experts from SSD, which our explainer on Janas covers.

What's confirmed, reported and uncertain

• Confirmed (Apple): the M5 Ultra specs, 1.2TB/s bandwidth, the 36-core CPU and 80-core GPU requirement for 512GB, the $5,499 starting price, "late October" timing for 512GB, and the Muse Spark Open, Qwen 3.8 and Mistral Medium 3.5 measurements in Apple's October 15 post (vendor-measured).
• Reported (MacRumors, Cult of Mac): that the 512GB option went on sale on October 28 and starts at $22,299 in the US, and Cult of Mac's $4,000 and $11,500 memory upgrade prices.
• Confirmed (Apple's store): the $6,799 starting price of the 36-core CPU, 80-core GPU model.
• Vendor-reported: the 4.3x, 4x, 9.8x and 3x performance claims.
• Illustrative calculations: every 4-bit estimate, the two-thirds budget and the tokens-per-second ceilings.
• Uncertain: Apple's own price for the 512GB configuration, ship dates, how well Mac runtimes handle NVFP4 and FP4 weights for these models, and Kimi K3.1's final weight format.