Mac LLM tokens per second: M5 Pro and M6 Mac minis, M5 Max and M1-M4

Quick answer

• $0: the Mac you own. An 8B model needs about 4.5GB at 4-bit. Check what yours runs.
• 16GB for 7B-14B. Used M4 mini 16GB/256GB, $749 median ask (Sept. 2): 24.11 t/s at 7B, measured. New M6, $899 (Apple, Sept. 2): about 30-31, estimate. Used M4 listings.
• M5 Pro Mac mini: 24GB for 14B, 48GB for 27B, 64GB for 70B. $1,699, $2,299 or $2,699 (Apple, Aug. 25). A 70B model on 64GB: about 6-7 t/s, estimate. Choose the tier or 24GB/512GB on Amazon.
• M5 Max Mac Studio for faster 27B or 70B. 36GB, $2,499 (Apple, Aug. 25): about 22-24 t/s on 27B, estimate. Used M4 Max 36GB/512GB, $2,299 median ask (Sept. 16): about 17-21, estimate. 64GB with the 40-core GPU, about $3,500 (Daring Fireball, rounded): about 11-13 on 70B. Apple.
• 96GB and up for 120B-class and larger mixture-of-experts. M5 Ultra 96GB, $5,499 (Apple, Aug. 25); gpt-oss-120B measured 79.68 t/s on an M2 Ultra. Routes for every tier.

What to buy for the model you want to run

Two numbers decide whether a model runs: the weights have to fit, and your context needs room on top. A third, memory bandwidth, decides how fast. The tool takes your budget and what you already own and answers keep, add a card or buy a machine, with the dated price and the retailer in the same view.

Model	Cheapest upgrade to a PC you own	Cheapest whole machine
The machine you already own — $0. Check this first. If it already has the “to work” figure free, keep it and spend nothing. The tool asks what you own before it shops and names what you would otherwise have paid.
Llama 3.1 8B4.5GB weights · 7.5GB to work	RTX 5060 Ti 16GB, new$790 · 16GB · 448GB/s	Mac mini M4, 16GB / 256GB, used$749 · 10.7GB · 120GB/s
Qwen 3.5 27B15.5GB weights · 25.9GB to work	Two used RTX 3090s$2,064 · 48GB · 936GB/s per card, not pooled	Mac Studio M5 Max, 36GB, new$2,499 · 27GB · 460GB/s
Llama 3.3 70B39.1GB weights · 65.2GB to work	No route we can source	GMKtec EVO-X2, Ryzen AI Max+ 395, 128GB$3,499 · 128GB · 256GB/s
gpt-oss-120b65.4GB weights · 109GB to work	No route we can source	GMKtec EVO-X2, Ryzen AI Max+ 395, 128GB$3,499 · 128GB · 256GB/s

Cheapest route that clears the recommended working figure, ranked the way the tool ranks it: on the total you pay — the dated price plus anything the price leaves out — with anything we found out of stock ranked below what you can order today. A card needs a desktop PC, a supply and the clearance to take it; a whole machine does not. The first answer is always to check what you already own — the tool asks that first, and says keep it when it fits.

The exact products behind those two figures

• RTX 5060 Ti 16GB, new$790 · lowest new listing we quote (Newegg), Sept. 15, 2026

  • ASUS Dual RTX 5060 Ti 16GB GDDR7 OC — 2.5 slotsAmazon · exact productOne specific board, not the listing the dated price above came from. No price or stock is shown here for it; check both on the listing.

• GMKtec EVO-X2, Ryzen AI Max+ 395, 128GB$3,499 · maker's own store price, 128GB + 1TB, Sept. 16, 2026

  • GMKtec's own store — the $3,499 128GB + 1TB configurationGMKtec · exact product · where the price above was readOpens on 128GB RAM + 1TB SSD / USThis is the price and stock state we read ourselves.
  • GMKtec EVO-X2, 128GB LPDDR5X, 2TB SSDAmazon · exact productThe 2TB configuration, not the 1TB priced above: the difference is storage, not memory for a model. No price or stock is shown here for this listing.

Other configurations and the option where you buy nothing are on the tool.

Answer it for my model and budget

Shopping during Prime Big Deal Days 2026? Match the model, context and CUDA requirement before comparing GPU, mini PC and Mac offers. The comparison keeps exact capacities, ordinary retailer asks and delivery limits separate from measured tokens per second.

For a separate memory-budget example, the Qwen3-8B context worksheet shows how prompt length and parallel requests add a 16-bit KV-cache payload to the weights, runtime allowance and other users of one memory pool. Its architecture-specific arithmetic does not apply to the other models in this article and does not predict tokens per second; keep the runtime, quantization and context of each benchmark below in view.

On the machines you can buy today, measured: a 7B model at 4-bit (Q4_0, llama.cpp) generates about 66 tokens a second on an M5 Pro with the 20-core GPU, 120 on a 40-core M5 Max and 92-94 on an M2 or M3 Ultra. gpt-oss-120B runs about 80 on an M2 Ultra. DeepSeek R1 671B at 4-bit runs 16-18 on a 512GB M3 Ultra in benchmark runs, and 4 to 11 with an 11,000-token prompt. The M5 figures come from MacBook Pros. The new M5 Mac mini and Mac Studio and every M6 have no benchmarks yet; their figures below are labeled estimates.

Choose a Mac by the model you want to run

Start with the largest model you will actually run. Its memory need sets the tier; the Mac options and a speed follow. Every speed names the chip it ran on and whether it was measured or estimated.

Minimum memory figures are canirun.ai estimates at roughly 4-bit quantization, not our measurements. Buy headroom: the operating system, other apps and a long conversation all need memory beyond the model file.

If no Mac in your range fits the model you want, the budget ladder starts at $0 and names the rung to look at instead, and the used-GPU tiers price 16GB, 24GB and 32GB cards with measured speeds. Once a tier fits, where to buy names the routes for each one.

Model class	Memory to buy	Mac options and price (dated)	Generation speed	Measured or estimate
7B-14B, such as Llama 3.1 8B (4.5GB minimum) or Qwen3 14B (7.8GB minimum)	16GB	Used M4 16GB/256GB, $749 median ask and a $625 single-listing floor (price ladder, Sept. 2); used 16GB/512GB M2 Pro at a $949.99 median ask and M4 at $958.87 (comparison, Sept. 2); M6 16GB/256GB, $899 new (Apple, Sept. 2)	37.87 t/s on an M2 Pro (16-core GPU) and 24.11 on an M4, 7B Q4_0, llama.cpp; about 34 t/s on M6, bandwidth arithmetic, and roughly 30-31 at the 16GB configuration’s 153GB/s	Measured on M2 Pro and M4; M6 estimate
About 27B dense, such as Qwen 3.5 27B (15.5GB minimum)	36GB	M5 Max 36GB, $2,499 new (Apple, Aug. 25); used M4 Max 36GB/512GB, $2,299 median ask (Sept. 16)	About 22-24 t/s on the M5 Max 36GB (460GB/s) and 17-21 on a used M4 Max 36GB (410GB/s), scaled from measured 7B llama.cpp runs; derivations	Estimate
70B dense, such as Llama 3.3 70B (39.1GB minimum)	64GB	M5 Pro Mac mini 64GB, $2,699 new (Apple’s base plus its upgrade, Aug. 25); M5 Max with 40-core GPU and 64GB, about $3,500 new (Daring Fireball’s August rundown, rounded)	About 6-7 t/s on the M5 Pro mini and 11-13 on the M5 Max, scaled from each chip’s measured 7B llama.cpp runs	Estimate
About 120B mixture-of-experts, such as gpt-oss-120B (65.4GB minimum)	96GB	M5 Ultra 96GB, $5,499 new (Apple, Aug. 25); M5 Max 128GB, about $5,100 new (Daring Fireball, rounded); used M1 Ultra 128GB, $5,249 median ask (Aug. 18-25)	79.68 t/s on an M2 Ultra 192GB, llama.cpp	Measured
About 235B mixture-of-experts, such as Qwen3-235B-A22B (131.3GB minimum)	256GB	M5 Ultra 256GB, $10,799 new (AppleInsider, Sept. 15; 256GB requires the 36-core CPU and 80-core GPU); used M3 Ultra 256GB, $9,999 median ask (Aug. 18-25)	16 t/s as GGUF and 24 under MLX on a 512GB M3 Ultra, LM Studio, per MacStories	Measured
About 671B mixture-of-experts, such as DeepSeek R1 (375GB minimum)	512GB	Used M3 Ultra 512GB, $17,448 median ask (Aug. 18-25); M5 Ultra 512GB, price not checked (Apple listed it as coming late October on Aug. 25)	18.11 t/s under MLX and 16.08 in a second 4-bit run on a 512GB M3 Ultra, per Hostbor	Measured; the second run’s runtime is not stated

For 7B-14B models, 16GB is the entry tier and more memory buys headroom. canirun.ai’s recommended figure for Qwen3 14B is 13GB, which leaves little of a 16GB Mac for the operating system and other apps.

Among 16GB/512GB used minis, generation follows bandwidth rather than chip age. In the llama.cpp table below, the base M2 Pro (16-core GPU, 200GB/s) generated 37.87 t/s on the 7B reference model and the M4 (120GB/s) generated 24.11. That is about 1.57x the M4’s speed at a Sept. 2 median ask $8.88 lower. The M4 remains the faster general-purpose Mac; the M2 Pro, M4 and M6 comparison covers CPU speed, ports and the new-machine premium. With 16GB, either machine is limited to small models, and long prompts cut speed on every row.

If 256GB of storage is enough, a used 16GB M4 at its $749 median ask costs $150 less than the $899 M6. The price ladder shows where a listing falls; its $625 floor is a single listing. The M6’s ~34 t/s comes from the estimate table below, not a measurement, and that table lists 153GB/s for the 16GB M6 configuration, so that machine may land a little lower: roughly 30-31 t/s by the same scaling. The M4 Pro measured 49.64 t/s at 273GB/s but is not priced in this piece.

The 36GB base M5 Max has a 32-core GPU and 460GB/s of bandwidth, per Apple’s specifications. The 614GB/s figure belongs to the 40-core configuration, which Daring Fireball’s rundown lists as the one with 64GB and 128GB upgrades. The used 36GB M4 Max is also the lower-bandwidth version: a 32-core GPU and 410GB/s, per Apple’s Mac Studio (2025) specifications. The chip table below lists both M4 Max configurations; the 40-core one starts at 48GB. For a 27B model, the used 36GB/512GB M4 Max is the cheaper route, at a Sept. 16 median ask $200 below the new M5 Max, and the M5 Max is the faster one, with 12% more bandwidth, if you can wait for Apple’s delivery. canirun.ai’s higher recommended figures — 65.2GB for Llama 3.3 70B, 109GB for gpt-oss-120B and 624.9GB for DeepSeek R1 — exceed those tiers. If you want room for long prompts or other apps, read each tier as a floor.

A 27B model does not require a Mac Studio. Its 15.5GB minimum also fits a 32GB M6 mini or a 48GB M5 Pro mini, with less bandwidth. Apple’s 16GB-to-32GB upgrade adds $400 to the M6, whose 16GB/512GB configuration lists at $1,099, and its 24GB-to-48GB upgrade adds $600 to the M5 Pro, which starts at $1,699 with 24GB and 512GB (Aug. 25, per our price-per-gigabyte table). By the same methods, those minis would generate roughly 8-10 and 16-19 t/s on a 27B model, against 22-24 on the 36GB M5 Max. The estimates below show each derivation, and the M5 Pro memory decision weighs each mini tier against a used RTX 3090, a 128GB mini PC and the Studio.

The 120B row’s speed was measured on an M2 Ultra, which is not one of the machines priced in that row. The M1 Ultra has the same 800GB/s bandwidth but generated 7B tokens about 11% slower in the chip table below. The M5 Max 128GB has about half the M5 Ultra’s bandwidth; we have not estimated its speed on this model.

No single Mac holds a model that needs more than 512GB; the cluster math prices that path. If you need frontier-model quality, the open models here are a tier below the $200 subscriptions; the subscription comparison weighs that trade. Long prompts cut speed on every row: in a separate DeepSeek R1 run, an 11,000-token prompt brought generation to 4.26 t/s under llama.cpp and 11.15 under MLX.

Measured Mac token-generation speeds by chip

The community-maintained llama.cpp Apple Silicon benchmark runs one model — LLaMA 7B — across every chip, which makes it the cleanest chip-to-chip comparison that exists. PP512 is prompt processing; TG128 is token generation. The Q4_0 file is 3.79GB and the F16 file 13.5GB, so the F16 column shows each chip on a much bigger read per token. The M5 Pro and M5 Max rows are the chips MacBook Pros have carried since March. All measured:

Chip	Bandwidth	PP512 t/s, Q4_0	TG128 t/s, Q4_0	TG128 t/s, F16
M4 (10-core GPU)	120GB/s	221.29	24.11	7.43
M2 Pro (16-core GPU)	200GB/s	294.24	37.87	12.47
M4 Pro (16-core GPU)	273GB/s	364.06	49.64	17.19
M4 Pro (20-core GPU)	273GB/s	439.78	50.74	17.18
M5 Pro (20-core GPU)	307GB/s	1,620.64	66.33	21.55
M1 Max (32-core GPU)	400GB/s	530.06	61.19	23.03
M4 Max (32-core GPU)	410GB/s	713.93	69.95	24.29
M4 Max (40-core GPU)	546GB/s	885.68	83.06	31.64
M5 Max (40-core GPU)	614GB/s	3,219.99	119.92	37.11
M1 Ultra (64-core GPU)	800GB/s	1,030.04	83.73	37.01
M2 Ultra (76-core GPU)	800GB/s	1,238.48	94.27	41.02
M3 Ultra (80-core GPU)	819GB/s	1,471.24	92.14	39.78

Nobody buys a Mac Studio to run a 7B model. The table’s value is the ratios. Up to the M4, generation speed tracks the bandwidth column, not the GPU core count and not the chip generation. An M4 Max from 2024 and an M1 Ultra from 2022 generate at the same speed because 546GB/s and a sublinearly-used 800GB/s land in the same place. The Mac mini rows show the same pattern: the 2023 M2 Pro out-generates the 2024 M4 with 66.7% more bandwidth.

The M5 chips break that pattern on a small 4-bit model. With about 12% more bandwidth than their M4 counterparts, the M5 Pro generated 31% faster than the 20-core M4 Pro and the M5 Max 44% faster than the 40-core M4 Max, passing every Ultra. Their prompt processing rose about 3.6-3.7 times, and a 4-bit 7B model spends enough of each token on compute to share in that. On the F16 file, 3.6 times larger, the gains shrink to 25% and 17%, and the M2 and M3 Ultras lead again. The larger the model, the more bandwidth decides.

How fast do large LLMs run on a Mac?

For gpt-oss-120B, the llama.cpp gpt-oss benchmarks record an M2 Ultra (192GB, 76-core) at 79.68 t/s generation. Olares published a served, single-stream figure of 69.39 t/s on an M3 Ultra — slower because serving stacks add overhead, and worth keeping because it is closer to what a user of a running system sees.

The biggest model that fits in one Mac is the stress test that counts. DeepSeek R1 671B at 4-bit on a 512GB M3 Ultra generates 16-18 t/s in Hostbor’s benchmark runs — 16.08 in one run and 18.11 under MLX, with no prompt length stated. MacRumors carried a similar 17-18 from Dave2D’s test when the machine launched. Feed it a real conversation and the story changes: at an 11,000-token prompt, one measured run got 4.26 t/s generation under llama.cpp and 11.15 under MLX, with the llama.cpp job taking about nine minutes end to end for the prompt plus a 1,662-token reply. Context depth, not model size, is where these machines actually hurt.

In between: Qwen3-235B-A22B runs about 16 t/s as GGUF and 24 t/s under MLX on a 512GB M3 Ultra at 4-bit in LM Studio, per MacStories’ testing. The runtime gap is general — MLX typically runs 10-30% faster than llama.cpp on the same hardware — so a tokens-per-second claim without a runtime label is missing a third of its meaning. Every figure in this piece names its runtime.

How can you estimate Mac tokens per second?

Token generation is memory-bandwidth-bound: each new token of a dense model requires reading essentially the whole model from memory. TensorFoundry’s roofline analysis of Apple Silicon puts real-world generation at about 75% of bandwidth divided by model size at 4-bit and about 87% at 8-bit. As a rule of thumb: bandwidth divided by model size in gigabytes, then take about three-quarters.

That rule assumes a dense model — in TensorFoundry’s words, “for each token, the model reads essentially all of its weights out of memory.” Mixture-of-experts models activate only part of theirs. Per their model cards, gpt-oss-120B has 117B parameters with 5.1B active, Qwen3-235B-A22B activates 22B of 235B and DeepSeek R1 activates 37B of 671B. For those models, total size decides whether the model fits, and the dense rule is a poor speed guide: applied to gpt-oss-120B’s 59.02GiB file on an 800GB/s M2 Ultra, it predicts about 9.5 t/s, and llama.cpp measured 79.68. For a mixture-of-experts model, use a measured figure for that model rather than either rule.

The rule degrades at the small end. Doubling bandwidth from M1 Max to M1 Ultra moved 7B generation from 61.19 to 83.73 — 1.37x, our arithmetic on the measured table — because small quantized models spend compute dequantizing instead of saturating the memory bus. The llama.cpp maintainer’s framing: quantized kernels “require extra compute to dequantize the data compared to F16 models.” Larger and less-quantized models get closer to the full bandwidth ratio.

Prompt processing is a different machine. It is compute-bound — “at large batch size the computation is compute bound,” per the same maintainer — and an MLX discussion of prefill performance gives the arithmetic: time to first token scales with roughly 2 x parameters x context length over available TFLOPS, and quantization does not help, because prefill reads the weights once and computes against every token. That is how a machine that generates 16 t/s spends two and a half minutes silent before the first token of a reply to an 11,000-token prompt.

This split is why Apple’s launch slides need translation. The claimed 3.9x (M5 Max) and 4x (M5 Ultra) gains are for “LLM prompt processing in LM Studio” — vendor-claimed, prompt-side numbers. llama.cpp’s measurements land near them: the M5 Max processed a 7B prompt at 3.6 times the 40-core M4 Max’s speed, and the M5 Pro at 3.7 times the 20-core M4 Pro’s. Generation gained far less, and less as the model grows: 31-44% on the 4-bit 7B model and 17-25% at F16, against official bandwidth gains of about 12% on both chips, 273 to 307GB/s and 546 to 614GB/s. The Ultra goes from 819GB/s to 1.2TB/s, up 50%, and has no measurement. The 32-core base M5 Max lists 460GB/s.

Estimated M5 and M6 Mac speeds, not measured

No figure below is a measurement of that machine at that model size. These are short-context estimates. The M5 Pro, M5 Max and used M4 Max rows scale the same chips’ measured runs in the table above; the M6 and M5 Ultra rows scale a predecessor’s measured speed by official bandwidth or apply the roofline rule. The M6 and M5 Pro Mac minis ship Sept. 22. Apple’s Aug. 25 delivery quotes for a New Jersey ZIP code put the base M5 Max Mac Studio by Oct. 5 and the base M5 Ultra by Oct. 12; the 512GB M5 Ultra had no price (delivery comparison). AppleInsider reported Sept. 15 that M5 Ultra deliveries run from Nov. 9 into January, depending on configuration. Measured numbers will arrive machine by machine, and we will add each one beside its estimate.

Machine	Bandwidth	Model class	Estimate, not measured	Basis
M6 Mac mini, 16GB	153GB/s	7B Q4	~30-31 t/s	M4’s measured 24.11 × 153/120
M6 Mac mini, 24GB or 32GB	170GB/s	7B Q4	~34 t/s	M4’s measured 24.11 × 170/120
M6 Mac mini, 24GB or 32GB	170GB/s	About 14B dense Q4	~15-20 t/s	170 ÷ 7.8GB × 0.7-0.9
M6 Mac mini, 32GB	170GB/s	About 27B dense Q4	~8-10 t/s	170 ÷ 15.5GB × 0.7-0.9
M5 Pro Mac mini, standard 16-core GPU	307GB/s	7B Q4_0	~65 t/s	Same chip’s measured 66.33 on the 20-core GPU, less the M4 Pro’s 2% 16-core gap
M5 Pro Mac mini, 18-core CPU, 20-core GPU	307GB/s	7B Q4_0	~66 t/s	Same chip configuration measured 66.33
M5 Pro Mac mini, 48GB or 64GB	307GB/s	About 27B dense Q4	~16-19 t/s	66.33 × 3.79GB to 21.55 × 13.5GB, ÷ 15.5GB
M5 Pro Mac mini, 64GB	307GB/s	70B dense Q4	~6-7 t/s	66.33 × 3.79GB to 21.55 × 13.5GB, ÷ ~40GB
Used M4 Max Mac Studio, 32-core GPU, 36GB	410GB/s	About 27B dense Q4	~17-21 t/s	69.95 × 3.79GB to 24.29 × 13.5GB, ÷ 15.5GB
M5 Max Mac Studio, 32-core GPU, 36GB	460GB/s	7B Q4_0	~90 t/s	40-core M5 Max’s measured 119.92 × 460/614
M5 Max Mac Studio, 32-core GPU, 36GB	460GB/s	About 27B dense Q4	~22-24 t/s	89.8 × 3.79GB to 27.8 × 13.5GB, ÷ 15.5GB; 27.8 is the 40-core F16 37.11 × 460/614
M5 Max Mac Studio, 40-core GPU	614GB/s	7B Q4_0	~120 t/s	Same chip configuration measured 119.92
M5 Max Mac Studio, 40-core GPU, 64GB	614GB/s	70B Q4	~11-13 t/s	119.92 × 3.79GB to 37.11 × 13.5GB, ÷ ~40GB
M5 Ultra Mac Studio	1.2TB/s	gpt-oss-120B	~104 t/s	M3 Ultra’s served 69.39 × 1.5
M5 Ultra Mac Studio	1.2TB/s	Qwen3-235B, MLX	~36 t/s	M3 Ultra’s 24 under MLX × 1.5
M5 Ultra Mac Studio	1.2TB/s	DeepSeek 671B Q4, MLX	~24-27 t/s	M3 Ultra’s 16-18 × 1.5

Before trusting a launch-week number

1. Runtime. MLX typically runs 10-30% faster than llama.cpp on the same hardware. A figure without a runtime label can’t be compared with these.
2. Quantization. The llama.cpp reference table uses Q4_0, which differs from the quants people actually run by 10-30% either way.
3. Prompt length. Short benchmark prompts flatter every machine. DeepSeek R1 went from 16-18 t/s to 4-11 at an 11,000-token prompt.
4. The exact configuration. The 16GB M6 lists 153GB/s rather than 170GB/s, the 32-core M5 Max lists 460GB/s rather than 614GB/s, and the measured M5 Pro has the 20-core GPU, not the standard mini’s 16-core. A result from one configuration does not describe the other.
5. Prompt processing or generation. Apple’s 3.9x and 4x claims are prompt processing. Generation, the reply speed, is what these tables estimate.

How the estimates were derived

The M5 Pro and M5 Max rows start from those chips’ measured runs. The 18-core M5 Pro mini has the measured configuration, a 20-core GPU. The standard mini’s 16-core GPU has the same 307GB/s; on the M4 Pro, the 16-core GPU generated about 2% slower than the 20-core, so its 7B row takes 2% off. No 32-core M5 Max has a measured row, so its 7B figure scales the 40-core chip’s 119.92 by 460/614.

For larger dense models, two measured runs set each range. The benchmark files are 3.79GB at Q4_0 and 13.5GB at F16, per ggml-org’s listing. Speed times file size gives the gigabytes a chip actually read per second; dividing by the larger model’s size gives its speed. The 4-bit run sets the low end, because a small quantized model spends more of each token on compute, and the F16 run sets the high end. The used 36GB M4 Max uses its own measured 32-core runs.

The M6 rows multiply the M4’s measured 24.11 by the bandwidth ratio, and its 14B and 27B rows are pure roofline — bandwidth divided by model size, times 0.7-0.9 — because no M6 has been measured. The M5 Ultra rows scale M3 Ultra measurements by 1.5, the ratio of 1.2TB/s to 819GB/s — the gpt-oss row from Olares’ served 69.39, the Qwen3 row from MacStories’ 24 under MLX, the DeepSeek row from the 16-18 benchmark range. Scaling the M2 Ultra’s llama-bench 79.68 instead would put gpt-oss near 120; serving overhead is real, so we quote the lower basis. The M5 Pro and M5 Max outran bandwidth scaling from their M4 predecessors, so the bandwidth-only M6 and M5 Ultra rows may prove low.

Model size uses canirun.ai’s minimum memory as a stand-in: 7.8GB for Qwen3 14B, 15.5GB for Qwen 3.5 27B and about 40GB for Llama 3.3 70B’s 39.1GB. A quantized file larger than that minimum would lower the estimate, and the roofline rule degrades at the small end, so the M6 14B row carries the most uncertainty. We left out 27B on a 24GB M6 or M5 Pro and 70B on a 48GB M5 Pro: each leaves less than 9GB beyond the model’s minimum for the operating system, apps and context. All of these are short-context figures; the DeepSeek context-depth collapse above applies to every row in proportion.

For a 70B model, the cheaper Mac trades away about half the speed. An 18-core M5 Pro mini with 64GB and 1TB lists at $3,199 (Apple, Aug. 25), about $300 under Daring Fireball’s rounded $3,500 for a 40-core M5 Max with 64GB, whose storage may differ. The mini’s chip has half the bandwidth, 307GB/s against 614GB/s, and measured 55% of the M5 Max’s 7B speed; scaled to 70B, that is roughly 6-7 t/s against 11-13.

Which Mac should you choose for local AI?

Start with the model-first table: find the largest model you will run, buy the memory tier above its minimum and judge speed by a measured figure for that model. For a new M6, M5 Pro or M5 Max, the estimates by configuration scale measured chips and bandwidth until each desktop is measured. For 7B-14B models on a Mac mini budget, the used M4 price ladder shows what a 16GB machine should cost. Price per gigabyte and bandwidth across Macs, Strix Halo and Nvidia are in the memory chooser and table, and what local capacity is worth against a subscription is in the fit comparison. Above 512GB, the cluster math covers what stacking Studios does and does not buy. Once the tier is settled, where to buy names the routes for each one, starting with the machine you already own.

Where to buy the machine your tier names

Prices below carry the source and date from the tier table. Match the chip, memory and storage on the listing, and confirm the current price at the store. Used Mac Studio asks dated Aug. 18-25 were recorded before most outgoing-model asks fell; in our matched panel, M3 Ultra asks were down 14.6% by Sept. 15, so open the linked listings before weighing a used Ultra against a new M5.

Tier and model class	What we recommend, dated	Route
$0. Whatever already fits the machine in front of you	Run the model on the computer you own. It costs nothing, and it shows whether memory or speed is the real limit before you spend.	Check what your machine runs
16GB. 7B-14B models	Used M4 mini 16GB/256GB, $749 median ask with a $625 single-listing floor (Sept. 2); new M6 16GB/256GB, $899 Apple list (Sept. 2)	Used: Swappa, eBay search, Apple certified refurbished or our M4 listings. New M6: Apple. Amazon: M4 16GB/256GB, Renewed Premium
24-64GB M5 Pro Mac mini. 14B, 27B or 70B dense models	24GB/512GB, $1,699; 48GB, $2,299; 64GB, $2,699, or $3,199 with the 18-core chip and 1TB (Apple, Aug. 25). Apple lists it available Sept. 22; custom configurations may take longer. Memory cannot be added later, and 48GB does not hold a 70B model with room.	New: Apple’s M5 Pro configurator. Which tier: the M5 Pro memory decision. Amazon: M5 Pro with 15-core CPU, 16-core GPU, 24GB/512GB
36GB. About 27B dense models	M5 Max Mac Studio 36GB, $2,499 new (Apple, Aug. 25); used M4 Max 36GB/512GB, $2,299 median ask (Sept. 16)	New: Apple. Used: our M4 Max listings or eBay search. Cheaper answer at this model size: a used 24GB graphics card. Amazon: M5 Max 36GB, 32-core GPU, 512GB
64GB. 70B dense models	M5 Max Mac Studio, 40-core GPU and 64GB, about $3,500 new (Daring Fireball’s August rundown, rounded); the 64GB M5 Pro mini above costs less and runs about half as fast	New: Apple. Cheaper answers: two used 24GB cards or a 128GB Ryzen AI Max+ 395 mini PC. Amazon: M4 Max 64GB, 40-core GPU, 1TB, last year’s chip at 546GB/s rather than the M5 Max’s 614GB/s
96-128GB. About 120B mixture-of-experts models	M5 Ultra 96GB, $5,499 new (Apple, Aug. 25); M5 Max 128GB, about $5,100 new (Daring Fireball, rounded); used M1 Ultra 128GB, $5,249 median ask (Aug. 18-25)	New: Apple. Used M1 Ultra: our listings or eBay search. Cheaper: a 128GB mini PC. Amazon: M4 Max 128GB, 40-core GPU, 512GB, a chip this row does not price, at 546GB/s against the M1 Ultra’s 800GB/s
256GB. About 235B mixture-of-experts models	M5 Ultra 256GB with the required 36-core CPU and 80-core GPU, $10,799 new (AppleInsider, Sept. 15); used M3 Ultra 256GB, $9,999 median ask (Aug. 18-25). Used, if price decides it: about $800 less; new buys 1.2TB/s against 819GB/s and a warranty. AppleInsider reported Sept. 15 that M5 Ultra deliveries run from Nov. 9 into January, depending on configuration.	New: Apple. Used: our M3 Ultra listings or eBay search. Amazon: M3 Ultra 256GB, 80-core GPU, 1TB
512GB. About 671B mixture-of-experts models	Used M3 Ultra 512GB, $17,448 median ask (Aug. 18-25); M5 Ultra 512GB, no Apple price as of Aug. 25	Used: our M3 Ultra listings or eBay search. New: Apple, which listed the 512GB M5 Ultra as coming late October with no price on Aug. 25. Amazon: M3 Ultra 512GB, 80-core GPU, 1TB

Storage is not the tier; memory is. The listings above carry 256GB, 512GB and 1TB SSDs, and Apple’s storage upgrades are the most expensive gigabytes on the page, so confirm the storage on the listing itself and read what an external SSD costs instead before paying for more.

Other configurations of the same tiers on Amazon: M4 mini 16GB/512GB and M4 Max Mac Studio 36GB/512GB.

Buying used? What protects you differs by channel: the channel-by-channel guide sets out warranties and returns, and the used-Mac checklist covers what to verify before paying. On the 16GB row, the M4 price ladder shows where an asking price is a discount.