Apple refreshed the lineup on 25 August 2026: the Mac mini M6 (from $899, 153–170 GB/s), the Mac mini M5 Pro (from $1,699, 307 GB/s), and a new Mac Studio with M5 Max (from $2,499, up to 614 GB/s) and M5 Ultra (from $5,499, 1.2 TB/s; 512 GB not orderable yet). Buy as much unified memory as you can.
512 GB M5 Ultra: Apple says the 512 GB option is coming late October; it was not orderable and had no price as of 24 October 2026, so a 512 GB Mac Studio today means the M3 Ultra. Apple holds an event on 28 October, and this page will be updated after it. Picks: new M5 Pro runs make LFM2.5-24B-A2B the index’s 64 GB pick (community114 tok/s on an M5 Pro); at 24–32 GB the pick is still Qwen 3.6-35B-A3B, which is the index’s #1 again (Score 81, needs 24 GB). New: Recursive Flint, a 40B coding MoE that fits a 24 GB Mac (community92 tok/s on an M5 Pro).
Shipping since 22 September, and the first measurements are in: BGR measured Qwen3.8-27B at about 55 tok/s on an M5 Ultra (the index had estimated 57), and a community run on a 32 GB Mac mini M6 measured Qwen 3.6-35B-A3B at 64 tok/s. Every speed below carries its provenance mark and is read from the LLMCheck index. Corrected in this update: the MacBook Pro M5 Pro and M5 Max have the same 307 and 614 GB/s as the desktops (this page listed 273 and 600), and the “Mac Studio M4 Ultra” it listed never existed — Apple skipped the M4 Ultra.
● Lineup: 25 August 2026 · Figures: 24 October 2026 · Prices via Amazon
Already own a Mac? Check what it can run →
Four machines cover almost everyone. Picks follow RAM, bandwidth and the index’s speed figures — never affiliate commissions.
Pick the largest model class you care about. We'll show the one Mac to buy — and one alternative.
🧠 Answer 3 questions — the Mac Advisor picks your config →
Already know your Mac? See the best LLMs for your exact Mac →
Sorted by price. Memory bandwidthd and max RAM are shaded by value — greener is more. See the full model leaderboard and benchmark data.
| Mac | Chip | Max RAM | Bandwidthd | GPU cores | Qwen 3.6-35B-A3B | Largest tier | From $ | |
|---|---|---|---|---|---|---|---|---|
| Mac mini M4 | M4 | 32GB | 120 GB/s | 10 | estimated~50 tok/s | ~14B dense; 30B MoE at 24GB+ | $599 | Check Price on Amazon → |
| Mac mini M6 Apple's first 2nm chip; 153 GB/s on the 16GB model. |
M6 | 32GB | 153–170 GB/s | 12 | community64 tok/s | 27–35B at 32GB | $899 | Check Price on Amazon → |
| MacBook Air M4 | M4 | 32GB | 120 GB/s | 8–10 | estimated~50 tok/s | ~14B dense; 30B MoE at 24GB+ | $1,099 | Check Price on Amazon → |
| Mac mini M4 Pro Previous generation — worth it only discounted. |
M4 Pro | 64GB | 273 GB/s | 16–20 | estimated~74 tok/s | 70B at 64GB | $1,399 | Check Price on Amazon → |
| Mac mini M5 Pro | M5 Pro | 64GB | 307 GB/s | 16–20 | community79 tok/s | 70B at 64GB | $1,699 | Check Price on Amazon → |
| MacBook Pro M5 Pro | M5 Pro | 64GB | 307 GB/s | 16–20 | community79 tok/s | 70B at 64GB | $2,199 | Check Price on Amazon → |
| Mac Studio M4 Max Previous generation. |
M4 Max | 128GB | 410–546 GB/s | 32–40 | estimated~105 tok/s | 70B / GLM-4.5-Air | $1,999 | Check Price on Amazon → |
| Mac Studio M5 Max 460 GB/s on the 36GB model (32-core GPU). |
M5 Max | 128GB | 460–614 GB/s | 32–40 | sourced118 tok/s | DeepSeek V4 Flash 2-bit at 128GB | $2,499 | Check Price on Amazon → |
| MacBook Pro M5 Max | M5 Max | 128GB | 460–614 GB/s | 32–40 | sourced118 tok/s | DeepSeek V4 Flash 2-bit at 128GB | $3,599 | Check Price on Amazon → |
| Mac Studio M5 Ultra Apple says the 512GB option is coming late October; it was not orderable and had no price as of 24 October 2026. |
M5 Ultra | 512GB | 1,228 GB/s | 64–80 | community153 tok/s | GLM-5.3 / GLM 5.2 at 512GB | $5,499 | Check Price on Amazon → |
| Mac Studio M3 Ultra The 2025 flagship; price shown for 256GB. |
M3 Ultra | 512GB | 819 GB/s | 60–80 | community106 tok/s | GLM 5.2 at 512GB | $6,999 | Check Price on Amazon → |
Speeds are single-stream decode for Qwen 3.6-35B-A3B at 4-bit, the index’s #1 model, read from the LLMCheck index: a community-measured figure where one exists for that chip (hollow circle), otherwise an estimate from the published bandwidth model (hollow diamond). LLMCheck does not run its own measurements. Max RAM is each machine’s largest configuration; prices are approximate starting prices — check the retailer. Compare specific chips: M5 Max vs M4 Max · M5 Pro vs M5 Max.
RAM determines which models you can load at all. Find your tier, see exactly what runs, and what won't.
Runs: on a 16GB Mac mini M6 (153 GB/s): Gemma 4 12B estimated~17 tok/s, GPT-oss 20B estimated~41 tok/s and Mellum2.1-12B-A2.5B-Thinking estimated~55 tok/s, plus NVIDIA’s new Nemotron 3.5 Bolt E4B estimated~49 tok/s. Every pick by job: best local LLM for a 16GB Mac.
Can't run: 27B+ models at 4-bit — too little memory (only the 1-bit and ternary Bonsai 27B builds fit).
Check Price on Amazon →Runs: on a 24GB Mac mini M6 (170 GB/s): Qwen 3.6-35B-A3B community64 tok/s — the index’s 24GB pick, with little room left for long context — plus Gemma 4 26B-A4B community48 tok/s, Qwen 3 30B-A3B sourced62 tok/s and the new coding model Recursive Flint (its 3-bit build) estimated~71 tok/s.
Can't run comfortably: dense 27B+ — it fits with zero room for context.
Check Price on Amazon →Runs: on a 32GB Mac mini M6: Qwen 3.6-35B-A3B community64 tok/s, Gemma 4 26B-A4B community48 tok/s, KAT-Coder-V2.5 estimated~47 tok/s, Recursive Flint estimated~61 tok/s, and dense Qwen3.8-27B community9 tok/s. The sweet spot for most buyers.
Can't run: 70B models.
Check Price on Amazon →Runs: on an M5 Pro (307 GB/s): LFM2.5-24B-A2B community114 tok/s — the index’s 64GB pick since 24 October — Qwen 3.6-35B-A3B community79 tok/s, Qwen3-Coder-Next estimated~49 tok/s and Llama 3.3 70B estimated~6 tok/s, with room for long context.
Can't run: 120B-class MoE at 4-bit such as Qwen 3.5 122B-A10B (65 GB) or Nemotron 3.5 Super VL (72 GB) — that is the 96–128GB tier; GPT-oss 120B (62 GB) just fits.
Check Price on Amazon →Runs: on an M5 Max (614 GB/s): Qwen3.8-Flash-Next community46 tok/s — the index’s 128GB pick — DeepSeek V4 Flash at 2-bit community39 tok/s and GLM-4.5-Air estimated~27 tok/s. Comfortable for 24/7 agents.
Can't run: 4-bit frontier MoE such as MiMo-V2.6-Flash — that needs 192GB+.
Check Price on Amazon →Runs: on an M5 Ultra (1.2 TB/s): MiMo-V2.6-Flash community62 tok/s, DeepSeek V4 Flash at 4-bit estimated~45 tok/s, GLM-5.3-Flash at 4-bit estimated~35 tok/s and MiniMax M2.5 estimated~57 tok/s. The new MiniMax-M3.1-Flash fits at 3-bit through MLX only, text-only for now (community35 tok/s). An M3 Ultra (819 GB/s) runs the same models about a third slower.
Caveat: trillion-parameter models such as Kimi K3 and DeepSeek V4 Pro stay server-class.
Check Price on Amazon →Runs: GLM-5.3 at 4-bit estimated~17 tok/s and GLM 5.2 estimated~22 tok/s on an M5 Ultra, and Meta’s 816B Muse Spark Open at 2–3-bit estimated~31 tok/s. Apple says the 512GB option is coming late October; it was not orderable and had no price as of 24 October 2026. Until it ships, a 512GB Mac Studio means the M3 Ultra. The index’s pick at 512GB stays MiMo-V2.6-Flash, which already fits 192GB; what 512GB adds is the models above.
Caveat: capacity, not speed — the biggest models decode at roughly 15–30 tok/s (estimated) even here.
Mac Studio M3 Ultra (512GB) on Amazon →The largest open models — Kimi K3, Qwen3.8-2.4T, DeepSeek V4 Pro and Inkling — are server-class: they don't fit even a 512 GB Mac Studio. To run the full, unquantized model, rent a datacenter GPU by the minute on Vast.ai — often 5–6× cheaper than AWS or GCP, with H100s and B200s available on demand.
Vast.ai referral link — we may earn a small commission at no extra cost to you. It never influences our reviews or rankings.
We'd rather you buy once and buy right. These are the things that actually change which Mac you should get.
Unified memory is soldered and can never be upgraded — buy the most RAM you can afford; it's the #1 factor in which models you can run.
Apple has no affiliate program. Amazon sells the identical Macs; we may earn a commission at no extra cost to you. It never influences our rankings.
Apple Certified Refurbished and Amazon Renewed offer the same machines for less — worth checking before you buy new, and often enough to afford the next RAM tier.
Memory bandwidth (GB/s), not just RAM size, drives tok/s — a Max chip is far faster than a Pro at the same RAM.
Local models are 4–250GB each. External storage and fast I/O keep your model library out of your boot drive.
Models are 4–250GB each — a 4TB portable SSD holds a serious library without filling your Mac's internal drive.
Check Price on Amazon →Pair an NVMe drive with a Thunderbolt 5 enclosure for a fast external model library that loads weights quickly.
Check Price on Amazon →For a desktop Mac mini or Studio — add displays, storage, and Ethernet over a single Thunderbolt 5 connection.
Check Price on Amazon →As an Amazon Associate, LLMCheck earns from qualifying purchases. The links above are affiliate links — they cost you nothing extra and help keep the LLMCheck index free and ad-light. Affiliate relationships never influence our rankings or recommendations; see our methodology for exactly how we score Macs and models.