Mac Studio M5 Max vs Strix Halo: What Apple's September 2026 Refresh Changes for Local LLMs

Quick answer: buy the Mac for dense models, buy a Strix Halo box for capacity per dollar and MoE. A 128GB Mac Studio with the M5 Max delivers 614 GB/s of memory bandwidth against roughly 215 GB/s measured on a 128GB Strix Halo mini PC — about 2.9× — but the $2,499 headline is a 36GB machine. Getting to 128GB runs Apple's memory ladder (+$600 for the 40-core GPU with 48GB, +$400 to 64GB, +$2,000 to 128GB), landing at roughly $5,099–$5,499 against $3,449 for a 128GB Framework Desktop. That is $1,650–$2,050 more for 2.9× the bandwidth and the same 128GB. Per gigabyte the AMD box still wins at $26.95/GB versus $39.84–$42.96/GB; per gigabyte-per-second of bandwidth, Apple now wins for the first time.

What changed on 22 September 2026

Apple announced a full desktop refresh on 25 August 2026 and shipped it on 22 September, discontinuing the entire M4 desktop line the same day. If you had a Mac on a shortlist last week, that shortlist is now four dead links.

Specs and memory ceilings are from Apple's Mac Studio tech specs; Mac mini figures and dates from Macworld's 2026 Mac mini rundown. Note the asterisk on the top of the range: the 512GB M5 Ultra config does not ship until late October 2026 — 256GB is the current ceiling you can actually order.

And the four machines Apple killed, each of which still appears in shopping guides written last month:

• Mac Studio M4 Max — $2,499+, discontinued. Up to 128GB at up to 546 GB/s. Replaced by M5 Max from $2,499.
• Mac Studio M3 Ultra — $5,299, discontinued. 96GB at 819 GB/s by the end (the 256GB and 512GB configs were pulled during the DRAM squeeze). Replaced by M5 Ultra from $5,499.
• Mac mini M4 Pro — $1,599, discontinued. Up to 64GB at 273 GB/s. Replaced by M5 Pro from $1,699.
• Mac mini M4 (base) — $799, discontinued. Up to 24GB at 120 GB/s. Replaced by M6 from $899.

All four are secondary-market buys now. That is not automatically bad news: a discounted M4 Max at 546 GB/s is 89% of the M5 Max's bandwidth, and if a reseller is clearing one it may be the best bandwidth-per-dollar machine in this entire article. We compared that generation head-to-head in Mac Studio M4 Max vs Strix Halo, which remains the right page for anyone shopping the used market.

On the AMD side the equivalent event has not happened yet. Minisforum showed a Ryzen AI Max+ PRO 495 "Gorgon Halo" MS-S1 MAX-P495 at IFA 2026 in early September with 192GB of unified memory, but no price has been announced, so it cannot be compared to anything. We walked through that decision in the Gorgon Halo wait-or-buy analysis.

The 128GB question: what Apple's headline price hides

This is the most useful section on the page, because it is the number Apple's marketing does not put in front of you and almost no launch coverage walks all the way to.

$2,499 buys 36GB. That is a fine machine and a bad local-LLM machine — 36GB does not hold a 70B model at 4-bit with any context worth having. To get the 128GB that competes with a Strix Halo box you climb a ladder, and per Macworld's M5 Max review (Roman Loyola, 21 September 2026) it goes like this: the 128GB option only exists on the 18-core CPU / 40-core GPU part, and that step "technically costs $300 but actually adds $600 to the price when the cost of the 48GB RAM upgrade is factored in. It can further be upgraded to 64GB ($400) or 128GB ($2,000)."

Why two numbers. Macworld lists 64GB and 128GB as alternative upgrades from the 48GB tier, which computes to $5,099. Read cumulatively through the 64GB step it computes to $5,499. Sources we checked disagree on which reading is right, and Apple's own configurator is the only authority — but apple.com/shop/buy-mac/mac-studio does not render prices to automated readers, so we could not settle it at write time. We publish the bracket and say why rather than pick a number and sound confident. Verify at Apple before you buy.

Two anchors that bound the bracket from outside. Macworld's as-tested unit — 128GB with a 4TB SSD — is $6,899, the only complete Apple-configured 128GB price a reviewer has published. And a well-specified 48–64GB M5 Max sits around $3,100–$3,600, which is the configuration most professional buyers should actually consider if they are not running 70B-class models. The jump from 64GB to 128GB is $2,000 for 64GB of memory — about $31 per gigabyte, charged on soldered LPDDR5X you can never change. Hold that number; it comes back in the last section.

Against that ladder, the 128GB x86 field has no ladder at all — 128GB is the SKU:

Note the MS-S1 Max caveat: the price is real and verified, but the listing was sold out at last check. Do not plan a purchase around it without confirming stock.

The inversion: Apple now wins bandwidth per dollar

Here is the arithmetic that flips the advice this site has been giving, computed longhand so you can check it.

Capacity per dollar — Strix Halo wins, comfortably:

• Framework Desktop: $3,449 ÷ 128GB = $26.95 per GB
• Mac Studio M5 Max: $5,099 ÷ 128GB = $39.84 per GB; at $5,499, $42.96 per GB
• Apple charges 1.48×–1.59× as much per gigabyte

Bandwidth per dollar — Apple wins, and this is new:

• Mac Studio M5 Max: 614 GB/s ÷ $5,099 = 0.120 GB/s per dollar; at $5,499, 0.112
• Framework Desktop at measured ~215 GB/s: 215 ÷ $3,449 = 0.062 GB/s per dollar
• Framework Desktop at theoretical 256 GB/s: 256 ÷ $3,449 = 0.074 GB/s per dollar
• Apple buys 1.8–1.9× more bandwidth per dollar against the measured figure, 1.5–1.6× against the theoretical one

Why that second ratio is the one that decides a purchase: memory bandwidth, not compute, caps token generation once a model fits in memory. Generating one token requires reading every active parameter out of memory, so tokens per second is roughly bandwidth divided by active bytes per token. Capacity decides whether a model runs; bandwidth decides how fast. We unpacked the mechanics in the Strix Halo bandwidth deep-dive and the concept itself in unified memory vs VRAM.

For a year the honest answer on this site was "Apple is faster but you pay a bandwidth tax you can't justify." The M5 Max's 614 GB/s at a 128GB price that rose less, proportionally, than the mini-PC field's did during the DRAM squeeze has closed that gap. If your workload is bandwidth-bound, Apple is now the rational buy, and we would rather say so than defend a recommendation the arithmetic no longer supports.

Dense vs MoE: the architecture split that decides your purchase

The bandwidth argument is not uniform across models, and this is where the recommendation splits cleanly in two.

A dense model reads every parameter for every token. Llama 3.3 70B at 4-bit means roughly 40GB read per token, and there is no way around it — the machine with more bandwidth wins by the full ratio. A Mixture-of-Experts model routes each token through a small fraction of its parameters. Qwen3 30B-A3B activates about 3B of 30B parameters per token; GPT-OSS 120B is similarly sparse. Far fewer bytes are read per token, so the bandwidth-starved machine spends proportionally less of its time waiting on memory — and Apple's advantage shrinks. See MoE and sparse vs dense if those terms are new.

Our own benchmark table shows the split plainly. Every Strix Halo row below is a real published run, not an estimate:

Read the two ends of that table against each other. A Strix Halo box runs a 120B MoE model at 31–34 tok/s — genuinely pleasant, faster than you read. The same box runs a dense 70B at 5 tok/s, which is a machine you leave running while you do something else. That 6–7× gap is not a quirk; it is the entire dense/MoE distinction expressed in one column.

So the split is:

• Dense 70B-class work — Llama 3.3 70B, DeepSeek R1 Distill Llama 70B, long-prompt summarisation, fine-tuning — is exactly where 5 tok/s hurts and where 2.9× the bandwidth is worth $1,650–$2,050. Buy the Mac.
• MoE work — GPT-OSS 120B, Qwen3 30B-A3B, Qwen3 32B — already runs well on hardware that costs $1,650 less, and the Mac's bandwidth advantage buys you less of an improvement than the same ratio suggests. Buy the Strix Halo box and put the difference toward storage or a second machine.

Size the model before choosing the box. Our VRAM calculator and the 128GB model guide will tell you which column you are in before you spend anything, and the measured tokens/sec page has the full Strix Halo picture.

What we cannot tell you yet — and what everyone else is inventing

You will find pages today quoting confident M5 Max tokens-per-second figures. They are fabricated. The chip shipped on 22 September 2026. As of 27 September, no credible independent LLM benchmark of any M5-family chip exists — not a llama-bench sweep, not an MLX run with a stated measurement spec, nothing with a methodology you could reproduce. The numbers circulating trace back to each other and ultimately to nothing. We checked, and we are not laundering them into our own voice.

What we can do is state an estimate and show its derivation so you can discard it when real data lands.

The naive estimate. Scale by bandwidth: 614 ÷ 215 = 2.86×, so a dense 70B that runs at 5 tok/s on a Strix Halo box would run at about 14 tok/s on a 128GB M5 Max.

The estimate we would actually bet on. The naive one is too high, and we know that from the previous generation. Tom's Hardware measured the M4 Max at roughly 1.6× a Strix Halo machine on token generation despite a 546 ÷ 215 = 2.54× bandwidth advantage — real-world scaling came in at about 63% of the ratio. Apply the same discount, or equivalently step the measured M4 Max result up by the 614 ÷ 546 = 1.12× bandwidth gain between generations, and you get roughly 1.8× a Strix Halo box: a dense 70B at something like 9 tok/s, not 14.

Nine tokens per second on a dense 70B is a real improvement over five — it is the difference between "usable for interactive work" and "usable for batch work." It is not a category change, and it costs $1,650–$2,050.

What would falsify this. Three things, and any one of them should make you ignore the estimate above:

• MLX maturity outruns the hardware. Apple claims the M5 Pro delivers up to 4× faster LLM processing than its predecessor — a first-party figure with no published methodology, but if the M5 generation's neural accelerators genuinely change the inference path rather than just the bandwidth, generation could beat the bandwidth ratio rather than trail it.
• Thermals. Sustained generation is a different load from a 30-second benchmark. The Studio chassis is well cooled, but a published tg128 number and an hour of real work are not the same measurement.
• Quantisation asymmetry. MLX 4-bit and GGUF Q4_K_M are not the same model. Cross-platform tok/s comparisons that ignore this overstate whichever side got the friendlier quant.

We will replace the estimate with measured figures the moment a credible third party publishes them with a stated measurement spec. Until then, the absence of a number is the finding — and anyone printing one this week is guessing with more confidence than we are.

Prefill and software: where the platforms really diverge

Token generation gets the headlines. Prompt processing — prefill, the time before the first token appears — is where these platforms separate most, and it is compute-bound rather than bandwidth-bound, so none of the arithmetic above predicts it.

Look at the prefill column in our benchmark table. A Strix Halo box processes GPT-OSS 120B prompts at 339.87 tok/s; the DGX Spark does 1,723.07 tok/s on the same model at the same measurement spec — roughly 5×. That gap is not about memory at all; it is about GPU compute and how mature the kernels are. If you paste 30,000-token documents into a model all day, prefill is your actual bottleneck and nobody's bandwidth chart tells you about it. Our prefill and time-to-first-token analysis covers this properly.

Apple has historically been weak here too — the M3 Ultra's known failing was slow prefill on long contexts despite 819 GB/s. Whether the M5 generation fixes that is, again, unmeasured.

On software, the honest ranking for local inference in 2026: MLX on Apple is mature, fast, and well-supported by LM Studio and Ollama, but it is an Apple-only island. ROCm on Strix Halo has improved a great deal and llama.cpp via Vulkan or HIP is solid for inference, but step outside inference — custom kernels, most fine-tuning scripts, anything research-grade — and you are porting. Neither platform has CUDA. That absence is the whole reason the DGX Spark stays in this conversation at $4,699+ despite losing on both ratios, and the reason GB10 boxes are the right answer for anyone whose toolchain assumes NVIDIA. If you want that comparison directly, we have DGX Spark vs EVO-X2 side by side.

One more Apple-side note: the M5 Ultra supports four-way Thunderbolt 5 clustering, which is the Apple analogue of the two-box pooling we tested in clustering two mini PCs. It is interesting and it is not cheap.

Why the cheaper Macs cannot do this job

The most common shortcut in this decision is "just get a Mac mini." One number kills it: capacity gates before bandwidth ever matters.

A 70B model at 4-bit is roughly 40GB of weights before you allocate a single byte of KV cache. The M6's 32GB ceiling locks it out entirely — no bandwidth figure rescues a model that does not fit. The M5 Pro's 64GB technically fits 70B at 4-bit, but with the OS resident and a working context you are scraping the ceiling, and it will not touch a 120B MoE. We worked the context arithmetic in how much context 128GB actually buys, and the sizing question generally in how much unified memory you need.

So Apple's real entry price for 70B-class local work is the 128GB M5 Max at roughly $5,099–$5,499. There is no $1,699 shortcut.

The honest budget floor, for completeness: if you are running 7B–14B models and want an always-on agent host rather than a 70B machine, a GMKtec M6 Ultra at $569 or a Beelink SER8 at $799–$939 does that job for a tenth of the money, and neither this article nor Apple's lineup is relevant to you. Our tiered buyer's guide covers that end.

Why both platforms got expensive at once

It is tempting to read the M5 Max's $5,099-to-$5,499 128GB tier as Apple being Apple. It is not — or not only. Both sides of this comparison were repriced by the same event.

Engadget's Steve Dent (21 September 2026) reports that Apple was paying roughly $6 per GB for LPDDR5X earlier in 2026 and roughly $10–11 per GB by September. The knock-on: the M5 Max base rose $500 and the M5 Ultra base rose $1,500 against their predecessors, and a 256GB Ultra that would have been $7,499 at the M3 generation's launch now lists at $11,299 as tested.

That is the same LPDDR5X squeeze that moved Strix Halo boxes from roughly $1,499 at launch to $3,449–$4,349 today, and that pushed NVIDIA to raise the DGX Spark from $3,999 to $4,699 in February 2026 citing memory supply. We traced the whole chain in why local AI mini PCs got expensive in 2026.

Two practical consequences. First, neither vendor is gouging you specifically — the memory market is, and every unified-memory machine on the market carries it. Second, and more useful: because the surcharge is on soldered memory you cannot upgrade later, buying less memory now to save money is an expensive decision you make once. Apple's $2,000 step from 64GB to 128GB is about $31/GB against a spot cost of $10–11/GB. Painful — and still the only way to get there.

Which to buy

Two SKUs with asterisks. The Minisforum MS-S1 Max at $3,799 is good value on paper but its listing was sold out at last check. The HP Z2 Mini G1a ($5,349–$7,406) and ASUS Ascent GX10 ($6,449–$7,999) both exist at 128GB but are priced above the Mac with no bandwidth answer; buy them for ECC memory or vendor support, not for local inference economics. If you want more than 128GB on the x86 side, the 192GB vs 128GB analysis covers what is coming.

Bottom line

A 128GB Mac Studio M5 Max delivers roughly 2.9× the memory bandwidth of a 128GB Strix Halo mini PC — 614 GB/s versus about 215 GB/s measured — for roughly $1,650 to $2,050 more, which makes Apple the better buy for dense models and a Strix Halo box the better buy for capacity per dollar and MoE models.

That is a real change from what this site recommended a month ago, and it is worth being precise about what changed and what did not. Apple did not get cheaper per gigabyte — it is still 1.5× the price per GB and the $2,000 step to 128GB is brutal. What changed is that 614 GB/s at that price is finally more bandwidth per dollar than the mini-PC field offers, which matters enormously if you run dense models and barely at all if you run MoE.

And the thing nobody else will tell you this week: nobody has measured an M5 Max running an LLM yet. Our 1.8× estimate is derived from the previous generation's measured scaling, clearly labelled, and we will replace it the day real numbers land. If a page quotes you an exact M5 Max tokens-per-second figure today, it made it up.

Start with the VRAM calculator to find out whether you are a dense buyer or an MoE buyer — that single answer decides this purchase more than any price in this article. If it says MoE, the Framework Desktop at $3,449 is the pick and you keep $1,650. If it says dense 70B, Apple just earned the premium.