On October 27, Apple put the M6 into the 14-inch MacBook Pro, from $1,999, and the iMac, from $1,499, with preorders opening October 28 in 35 countries and regions and machines in stores on November 6. The chip is not new: Apple introduced it on August 25 in the Mac mini, calling it its "first state-of-the-art 2-nanometer chip." What makes this round worth an engineer's attention is the claim Apple keeps making for it, that these machines run LLMs locally, and the specs that decide whether that holds are not the ones printed on the box.
What shipped, and what was already shipping
The MacBook Pro keeps its chassis and gains the full chip: a 12-core CPU and 12-core GPU, up from 10 and 10 on the M5, and a dual 16-core Neural Engine (Apple). Memory starts at 16GB and stops at 32GB. Storage starts at 1TB, because Apple dropped the 512GB tier in March.
Against the M5 MacBook Pro, Apple claims up to 30% faster graphics, up to 25% faster on-device AI, a Neural Engine up to 1.9x faster and Xcode builds up to 1.3x faster, plus up to 24 hours of battery life, its longest in any Mac. These are Apple's numbers from Apple's bench. MacRumors adds that Wi-Fi 7 and Bluetooth 6 now run through Apple's N1 wireless chip, which Apple's release does not name.
The iMac's $1,499 buys a 10-core CPU, 10-core GPU version of the M6; the full chip starts at $1,699 (Apple). The cheaper model also caps memory at 24GB, leaving 32GB to the 12-core version, and both start at 256GB of storage (Fstoppers). Apple's iMac comparisons run against the M4 model, claiming up to 5x its on-device AI prompt processing in LM Studio. HotHardware identifies the process as TSMC's N2.
Three tiers, and fewer of the fastest cores
Apple's August release gives the layout: two super cores, four performance cores and six efficiency cores. The M5 paired four super cores with six efficiency cores (MacRumors); the M5 Ultra uses 12 super cores, 24 performance cores and no efficiency tier (Apple). Read across that lineup, the tiers sort by job: super cores for latency-bound single threads, where Apple claims the world's fastest CPU core; performance cores for parallel throughput; efficiency cores for the background hum. The M6 gives up two of the M5's fastest cores in exchange for four throughput cores, and Apple claims up to 1.2x the M5's multithreaded performance.
The only independent numbers come from the M6 Mac mini, which has shipped since September 22. A Geekbench 6.3 listing from a suspected pre-production unit scored 4,610 single-core and 20,676 multi-core, roughly 10% and 20% above the M5's average (Notebookcheck, in German). A Geekbench 7 result that surfaced via posts on X put multi-core at 22,783, level with the 14-core M3 Max of 2023 at 22,644 (Notebookcheck). Neither listing is verified, neither reflects a laptop's thermals, and neither says anything about token generation.
The arithmetic that decides what runs
Start with the fine print. Apple promises memory bandwidth of up to 170GB/s; MacRumors, citing Apple's spec sheet, reports that only the 24GB and 32GB configurations get it. The $1,999 model, at 16GB, runs at 153GB/s, exactly the M5's figure. On the bus that sets generation speed, the base M6 is an M5.
Now the division, which is our estimate. Generating each token from a dense model means streaming every weight across the memory bus, so bandwidth divided by model size caps decode speed. A 27-billion-parameter model quantized to about 4.5 bits per weight weighs roughly 15GB. At 170GB/s the ceiling is about 11 tokens per second; at 153GB/s, about 10.
The measurements, compiled by Starry Hope from YouTube testers, sit against that wall. Alex Ziskind got about 9.7 tokens per second from a 27B Qwen model on a 32GB M6 Mac mini, with about 23GB of the machine's memory in use. Execute Automation got 8 to 9 with 4-bit MLX, and 16 to 19 with multi-token prediction, which wrings more than one token from each pass over the weights. Ziskind's STREAM run measured 143 to 145GB/s, below Apple's rating. The twelfth CPU core appears nowhere in this math.
Apple's 25% figure belongs to the other half of inference; its own example is LLM prompt processing. Prefill is compute-bound, which is where the GPU's Neural Accelerators earn their keep: on a 4-bit 9B model in llama.cpp, Ziskind measured 742 tokens per second of prefill against an M4's 210, while generation rose only from 18 to 26.9 (Starry Hope). Compute sets how fast a model reads; bandwidth sets how fast it writes. Scaled by model size, that prefill rate suggests a 27B model needs around two minutes to ingest a 30,000-token slice of a codebase (a rough estimate; attention makes long contexts worse).
Capacity sets the outer wall. A 70B dense model at the same quantization needs about 39GB and does not load. Mixture-of-experts models bend the division: a 30B model with about 3B active parameters, the shape of Qwen3-30B-A3B, needs about 17GB resident but reads under 2GB per token, for a theoretical ceiling near 90 to 100 tokens per second. Real runs fall short of ceilings; the ratio is the point. Capacity decides what loads, and active parameters times bandwidth decides how fast it speaks.
Thirty-two gigabytes in an expensive year
Against the desk-side field, the M6 is the small machine. Nvidia's DGX Spark moves 273GB/s over a 256-bit LPDDR5X bus to 128GB (Tom's Hardware); AMD's Ryzen AI Max+ 395 manages about 256GB/s to the same 128GB (AIMultiple's figures, via Starry Hope). On the 27B division, those are ceilings near 18 and 17 tokens per second. The larger gap is capacity: both run a 120B-parameter model at roughly 34 to 38 tokens per second in The Register's tests (Starry Hope), a pace only a mixture-of-experts design allows at that bandwidth, and a model like OpenAI's gpt-oss-120b needs about 60GB (Tom's Hardware), nearly twice what any M6 can hold. Within Apple's ladder, the M5 Pro offers 307GB/s and up to 64GB (Starry Hope), the M5 Ultra 1.2TB/s and up to 512GB (Apple). Climbing the lineup is mostly a way of buying bandwidth.
None of that memory is cheap this year. In June, Apple raised the 1TB MacBook Pro from $1,699 to $1,999, citing memory and storage costs, as TrendForce put first-quarter DRAM price rises as high as 98% with another 58 to 63% expected in the second quarter (Reuters). The National reports that the M6 machines simply hold those post-increase prices. The M6 Mac mini rose $100, to $899 (Tom's Guide), and the Spark's list price went from $3,999 to $4,699 in February (Starry Hope). The memory that would let you run AI at your desk is scarce because AI is running everywhere else.
Above the M6, Bloomberg's Mark Gurman reported in June that Apple will skip M6 Pro and M6 Max, with a base M7 in the first half of 2027 and M7 Pro and Max at the end of that year (Engadget); that is a report, not an Apple announcement. The touchscreen OLED MacBook Pro shown October 28, covered in a separate post, stays on M5 Pro and M5 Max, which fits the report.
The 30,000-foot read
Strip away the core counts and the M6 MacBook Pro, in its 24GB and 32GB forms, is a modest and sensible inference machine: a 27B-class dense model at around 10 tokens per second, a mid-sized MoE model at a comfortable clip, all of it private and portable. The $1,999 configuration is not that machine. Sixteen gigabytes at 153GB/s is last year's local-AI ceiling with this year's processor bolted on, and the cores Apple counted for you are the part of the chip a language model barely touches while it writes.
The stoic division applies. Within an engineer's control: buying the bandwidth tier, choosing models by active parameters rather than headline size, and letting techniques like multi-token prediction recover what the bus cannot deliver. Outside it: the price of DRAM while datacenters bid for every wafer. Watch for independent tests of the laptop and iMac after November 6, and for whether the M7 arrives with a wider bus. That, not another efficiency core, is what moves the number that matters.
Sources: Apple Newsroom, 14-inch MacBook Pro with M6 (primary: prices, dates, 35 countries and regions, 16GB to 32GB memory, bandwidth of up to 170GB/s, performance and battery claims against M5, local LLM positioning); Apple Newsroom, iMac with M6 (primary: $1,499 10-core and $1,699 12-core models, M4 comparisons); Apple Newsroom, M6 and M5 Ultra (primary: August 25 introduction, 2nm wording, core tiers, 1.2x multithreaded claim, M5 Ultra cores, 1.2TB/s and 512GB); MacRumors (M5 core layout, N1 chip, 1TB base storage); MacRumors, bandwidth split (153GB/s on 16GB, 170GB/s on 24GB and 32GB, per Apple's tech specs); Daring Fireball (Mac mini launch and September 22 ship date; corroborates Apple's 2nm wording); Fstoppers (iMac memory caps and storage); HotHardware (TSMC N2); Notebookcheck, German (Geekbench 6.3 against M5); Notebookcheck (Geekbench 7 against M3 Max); Starry Hope, M6 Mac mini for local AI (third-party token rates, STREAM bandwidth, M5 Pro bandwidth and capacity); Starry Hope, DGX Spark landscape (Strix Halo bandwidth, The Register's 120B results, Spark pricing); Tom's Hardware, DGX Spark review (Spark bandwidth and memory) and page 3 (gpt-oss-120b memory, Qwen3-30B-A3B); Reuters via Khaleej Times, June 25 (Apple price increases, TrendForce DRAM figures); Reuters via Khaleej Times, October 27 (October 28 touchscreen MacBook Pro event); The National (prices held at post-increase levels); Tom's Guide (Mac mini $899, up $100); Engadget (Bloomberg's M6 Pro and Max report). CAVEATS: every performance multiplier is Apple's own claim. Geekbench results come from the M6 Mac mini, not these machines; the 6.3 listing is a suspected pre-production unit and the 7 result came via posts on X; neither is independently verified. Token rates are YouTube testers' measurements as compiled by Starry Hope; we did not review the videos. Decode ceilings, the prefill time and the 70B and MoE sizing are our estimates at an assumed 4.5 bits per weight, not measurements. Strix Halo bandwidth and The Register's 120B figures reach us third-hand via Starry Hope; Spark pricing is single-source. The M6 Pro and Max report is Bloomberg's, unconfirmed by Apple. The N1 chip is named by MacRumors, not by Apple's releases. Reuters dated Apple's price increase June 25; The National places it in July. iMac bandwidth by configuration and the 24GB and 32GB upgrade prices were not in the sources we could reach. TechCrunch and Ars Technica coverage did not surface in our searches and is not cited. Tom's Guide's description of the M6 as the world's first commercially available 2nm chip is unverified and not used. The October 28 machine's M5 Pro and M5 Max chips come from our separate coverage. Corrections to early framing: the M6 debuted August 25 in the Mac mini, not in these machines; its 2nm process is stated by Apple itself; the $1,499 iMac uses a 10-core CPU, 10-core GPU M6; and 170GB/s excludes the 16GB base MacBook Pro.