Mac Studio M5 Max and M5 Ultra for Local LLMs: Which Configuration to Buy

Buy the M5 Mac Studio whose unified memory holds the model file plus context, then use bandwidth and list price to break ties. As of Apple’s September 22, 2026 newsroom post, Mac Studio with M5 Max (from $2,499) and M5 Ultra (from $5,499) is on sale. The 512GB memory configuration is the one still scheduled for late October. LocalRig has not yet measured M5 tokens per second. Apple’s “faster AI” multiples are vendor prompt-processing tests from July 2026, and the only LocalRig Apple tok/s row is a base M4.

This page used to be a wait-or-buy rumor note. The URL is unchanged so older links still land here. The decision is now which configuration to order. Size the file in the VRAM calculator before you pick a tier. The calculator adds weights, a KV estimate, and 15 percent headroom. That recommended figure is the planning margin in the tool. It is not a measured macOS footprint.

Which memory tier matches the model?

Apple’s spec page (fetched 2026-10-06) lists two chips and fixed memory steps. Memory is chosen at purchase.

Configuration	Unified memory	Bandwidth	GPU	US list price, 2026-10-06	What already-published LocalRig pages put in this class
M5 Max, 18-core CPU	36GB	460GB/s	32-core	$2,499 (512GB SSD)	Above a 24GB card. Qwen 3.8 27B Q4 is that 24GB class. Long context is a separate bill.
M5 Max, 18-core CPU	48GB	614GB/s	40-core	$3,099 (512GB SSD)	70B Q4_K_M is ~35–40GB of weights and a 46–48GB comfortable floor with KV. 48GB unified is the edge of that floor once macOS is also resident.
M5 Max, 18-core CPU	64GB	614GB/s	40-core	$3,499 (512GB SSD)	Same 70B Q4 class with more spare than 48GB. Still outside the site’s 128GB lane.
M5 Max, 18-core CPU	128GB	614GB/s	40-core	$5,099 (512GB SSD); $5,399 with 1TB	Qwen 3.8 27B at 200k and DeepSeek-V4-Flash 3-bit ~103GB. MiniMax-M3 1-bit wants ~133GB — tight on 128GB, not a promise.
M5 Ultra, 30-core CPU	96GB	1.2TB/s	64-core	$5,499 (1TB SSD)	More bandwidth than any M5 Max, less memory than the 128GB Max. A 75GB Qwen 3.8 Flash GGUF is a large file inside 96GB before KV.
M5 Ultra, 36-core CPU	256GB	1.2TB/s	80-core	$10,799 (1TB SSD)	MiniMax-M3 3-bit at 164–200GB (Unsloth figures on that page, accessed 2026-09-10). Kimi K2.6 at a 350GB working rule does not fit.
M5 Ultra, 36-core CPU	512GB	1.2TB/s on the Ultra spec line	80-core	Price not listed. Apple: late October.	K2.6’s 350GB rule is the class a 512GB pool can hold. Kimi K3 1-bit at 594GB / ~610GB and DeepSeek-V4-Pro Q4 ~850GB still miss.

The 40-core M5 Max is the chip that offers 48GB, 64GB, and 128GB. The $2,499 machine is the 32-core GPU at 460GB/s and 36GB. Stepping to the 40-core GPU on Apple’s store requires the 48GB memory upgrade. Prices above are schema.org Offer amounts on the named US configuration pages, fetched 2026-10-06. Education pricing is lower (newsroom: M5 Max from $2,299, M5 Ultra from $5,099) and is not the table.

Run the calculator for your quant and context. A 128GB machine does not give you 128GB of weights. How much unified memory is the older tier essay; treat its M3/M4 bandwidth rows as that generation, and use the table here for M5 steps.

What does the bandwidth change, if tok/s are not measured?

Decode speed tracks memory bandwidth: every token re-reads the weights. On the spec sheet, M5 Max is 460GB/s (32-core GPU) or 614GB/s (40-core GPU). Both M5 Ultra GPU configs are 1.2TB/s. Apple’s August 25, 2026 release calls 1.2TB/s “50 percent higher than before.” The M3 Ultra page records the previous Studio generation at 800GB/s. This page does not turn that ratio into an M5 tok/s estimate.

M5 Max and M5 Ultra decode tokens per second: not yet measured. LocalRig’s own Apple bench is a base M4, 16GB, Llama 3.1 8B Q4_K_M, 18.4 tok/s on llama.cpp b9820 and 19.5 tok/s on Ollama 0.30.11, measured 2026-06-27. That is an 8B result on an M4. It is not an M5 Max or M5 Ultra result, and it is not a 70B result.

What Apple did publish, labeled here as vendor, from testing Apple says it ran in July 2026 (August 25 newsroom; see apple.com/mac-studio for their method):

• M5 Max, LLM prompt processing in LM Studio: up to 3.9× the prior-generation comparison they name as M4 Max, and up to 10.7× versus M1 Max.
• M5 Ultra, LLM prompt processing in LM Studio: up to 4× versus M3 Ultra, and up to 9.8× versus M1 Ultra.
• Peak AI compute: up to 4.3× M3 Ultra on M5 Ultra. M5 Max: up to 3.9× “the prior generation,” which that section ties to prompt processing.

Those sentences are about prompt processing and peak AI compute. They do not include a model name, quant, context length, or tokens per second. Why prompt processing is slow on Mac is still the mechanism page. Neural Accelerators in each GPU core are Apple’s description of the hardware behind those vendor multiples.

What do you pay per configured gigabyte?

“Per configured GB” divides the whole-computer list price by the unified-memory number on that configuration. It includes the chip and the SSD in the price. It is not a price for usable gigabytes after macOS, and it is not the marginal price of a memory upgrade.

Configuration (SSD in the price)	List price	$ per configured GB
M5 Max 36GB / 512GB SSD	$2,499	$69.42
M5 Max 48GB / 512GB SSD	$3,099	$64.56
M5 Max 64GB / 512GB SSD	$3,499	$54.67
M5 Max 128GB / 512GB SSD	$5,099	$39.84
M5 Ultra 96GB / 1TB SSD	$5,499	$57.28
M5 Ultra 256GB / 1TB SSD	$10,799	$42.18
M5 Ultra 512GB	not listed 2026-10-06	not computed

On the 40-core M5 Max with a 512GB SSD, the memory steps are a flat $25 per extra configured GB: 48GB to 64GB is +$400 for 16GB, and 64GB to 128GB is +$1,600 for 64GB. The 128GB Max is the cheap step on that ladder if your file needs it. The 96GB Ultra costs more per configured GB than the 128GB Max ($57.28 versus $39.84) because you are paying for 1.2TB/s and the Ultra CPU/GPU, and you get less memory.

The jump from the 96GB Ultra ($5,499, 30-core CPU, 64-core GPU, 1TB) to the 256GB Ultra ($10,799, 36-core CPU, 80-core GPU, 1TB) is +$5,300 for 160GB, or $33.13 per extra GB, and that step also changes the CPU and GPU. It is not a pure memory upgrade. A chip-picker “from $6,799” figure appears in the page script for the 36-core/80-core Ultra. The priced 256GB/1TB configuration itself is the $10,799 offer. Use $10,799.

What do power and noise look like under load?

LLM watts and LLM fan noise: not yet measured. These are Apple’s published figures for other tests.

Apple Support’s power table, published September 21, 2026, measures wall power with no external peripherals. “Max” is a compute-intensive test application, not an LLM run. “Idle” is Finder only, at about 20.2°C:

• M5 Max, 18-core CPU, 32-core GPU, 36GB, 512GB SSD: 7W idle, 200W max.
• M5 Ultra, 36-core CPU, 80-core GPU, 512GB memory, 16TB SSD: 9W idle, 385W max.

The spec sheet’s 480W “maximum continuous power” is the electrical rating of the machine, not that idle/max table. The August 2026 environmental report lists a different idle protocol (“Idle—Display on,” Ethernet connected): about 5.4W to 5.8W depending on line voltage. That is not the 7W/9W Finder figure. Quote the test you mean.

Acoustics on Apple’s UK M5 tech-spec page, fetched 2026-10-06: sound pressure at the operator position is 7 dB at idle and 7 dB on their wireless-web test (sound power 1.5 B, Kv = 0.3). The configuration tested is M5 Ultra, 30-core CPU, 64-core GPU, 96GB, 1TB. Wireless web is six browser tabs. It is not a saturated GPU inference load.

When is an M3 Ultra, or a non-Mac box, the better buy?

Keep the M3 Ultra you already have when the model fits and the open question is decode speed. M5 Ultra’s spec bandwidth is higher (1.2TB/s versus the 800GB/s on the M3 page), and Apple’s vendor table claims up to 4× M3 Ultra prompt processing in LM Studio. Neither of those is a LocalRig decode measurement. Replacing a working M3 to chase an unpublished tok/s number is a bet, not a result. The M3 page’s note that new M3 Studios lost the 512GB option is about that generation. M5 lists 512GB again, arriving late October, price still unpublished on the URL checked October 6.

A used or refurbished M3 Ultra is the better new-money purchase only when the price you can verify is lower, per configured GB, than the new M5 tier that actually holds your file. LocalRig did not re-survey used M3 Ultra listings on 2026-10-06. Do not reuse older “$10,000–$14,000” lines from the M3 article as today’s street price.

Buy a non-Mac box, or rent, in three cases:

• The weights fit in 24GB and you want CUDA, vLLM, or more than one user. Start with the used RTX 3090 guide and the best GPU guide. This page is not a fresh street-price check on those cards.
• You want a quiet desk machine and you are not on MLX or Metal. The existing comparisons are Strix Halo versus Mac Studio and DGX Spark versus two RTX 3090s. A separate 64GB-versus-128GB Spark article is still on an unmerged branch, so it is not linked here.
• The working set is past 512GB (Kimi K3, DeepSeek-V4-Pro Q4). One Studio does not hold it. See Mac Studio clusters and rent versus buy. Apple says a cluster of four Mac Studios delivered up to 3× the AI inference of one system in their July 2026 test. That is a vendor cluster claim, not a LocalRig packet, and it does not name the model.

Mac mini is a different ceiling. The same September 22 newsroom post puts Mac mini with M6 at $899 (16GB, configurable to 32GB) and M5 Pro at $1,699 (up to 64GB, 307GB/s on the spec page). If 64GB is enough, read Mac mini versus Mac Studio before paying Studio prices. The mini does not reach 128GB.

Who this is NOT for

• CUDA, vLLM, TensorRT-LLM, or a multi-user server. A Mac Studio is a single-user Metal/MLX machine. The bandwidth table does not make it a serving box.
• A 24GB model and a tokens-per-dollar constraint. The $2,499 M5 Max is 36GB at 460GB/s. A used 24GB NVIDIA card is the comparison, on the GPU guides, once you recheck that card’s price.
• Kimi K3 or DeepSeek-V4-Pro Q4 in one chassis. 512GB unified, when it ships, is still short of ~610GB and ~850GB working sets.
• Anyone who needs an M5 tok/s number before the order. That cell is empty. Waiting for LocalRig to measure it is a real choice. Pretending the M4 8B row fills it is not.

The configuration to order

If the calculator’s recommended figure fits in 128GB, order the M5 Max 40-core, 128GB, 512GB SSD at $5,099 (2026-10-06). It is the low point on dollars per configured GB among the priced M5 tiers, at 614GB/s. Step up to the M5 Ultra 36-core, 256GB, 1TB at $10,799 when the file is in the 128–256GB band and you want the 1.2TB/s pool. Take the 96GB Ultra at $5,499 only when you specifically want 1.2TB/s for a model that fits in 96GB and you have already rejected the 128GB Max’s extra memory. Leave 512GB until Apple posts a price. Leave M5 tok/s blank until someone publishes a model, quant, runtime, context, and hardware together.

M5 Max configurations on Amazon → · M5 Ultra configurations on Amazon →

Those Amazon links are affiliate search links (tag localrig-20) and may earn a commission. The prices in this guide are Apple’s US list prices, checked on apple.com on 2026-10-06: Mac Studio store.

Sources

• Apple Newsroom, September 22, 2026, “The new Mac mini and Mac Studio are available today.” Starting prices, on-sale date, 512GB in late October, 1.2TB/s, Thunderbolt 5 RDMA. https://www.apple.com/newsroom/2026/09/the-new-mac-mini-and-mac-studio-are-available-today/
• Apple Newsroom, August 25, 2026, Mac Studio with M5 Max and M5 Ultra. Vendor LM Studio prompt-processing multiples; testing conducted by Apple in July 2026. https://www.apple.com/newsroom/2026/08/apple-introduces-new-mac-studio-with-m5-max-and-m5-ultra/
• Apple, Mac Studio technical specifications, fetched October 6, 2026. Memory steps, 460GB/s, 614GB/s, 1.2TB/s, 480W maximum continuous power. https://www.apple.com/mac-studio/specs/
• Apple US online store JSON-LD Offer prices, fetched October 6, 2026, for the configuration URLs named in the tables (36GB $2,499, 48GB $3,099, 64GB $3,499, 128GB/512GB SSD $5,099, 128GB/1TB $5,399, 96GB $5,499, 256GB/1TB $10,799). The 512GB-memory URL returned no priced offer that day.
• Apple Support, “Mac Studio power consumption and thermal output,” published September 21, 2026. https://support.apple.com/en-us/102027
• Apple Support (UK), Mac Studio (M5 Max or M5 Ultra) tech specs, acoustic table, fetched October 6, 2026. https://support.apple.com/en-gb/128107
• Apple, Mac Studio Product Environmental Report (August 2026 PDF), idle-with-display-on power. https://www.apple.com/environment/pdf/products/desktops/Mac_Studio_PER_Aug2026.pdf
• LocalRig first-party, base M4 16GB only: llama.cpp b9820 at 18.4 tok/s and Ollama 0.30.11 at 19.5 tok/s, Llama 3.1 8B Q4_K_M, 2026-06-27. Not an M5 measurement.
• Model-size figures are the ones already published on the linked can-i-run pages (Unsloth and vendor cards, accessed 2026-09-10 on those pages). They are not new M5 benchmarks.