M6 MacBook Pro for AI: what the Dual 16-core Neural Engine, 32GB memory ceiling and 170GB/s bandwidth mean for running local models

Update, October 29, 2026: At the other end of Apple's range, MacRumors reported that the Mac Studio M5 Ultra with 512GB of unified memory went on sale on October 28, at a US price of $22,299. Apple lists 1.2TB/s of memory bandwidth for the M5 Ultra. See what 512GB means for the biggest open models.

How good is the M6 MacBook Pro for AI? Apple announced the 14-inch MacBook Pro with M6 and a new iMac with M6 on October 27, 2026, and pitched both on on-device AI: a new "Dual 16-core Neural Engine," Neural Accelerators in every GPU core and faster "LLM prompt processing." For running open models locally, though, the numbers that matter most are memory size and memory bandwidth. Both M6 machines top out at 32GB of unified memory and up to 170GB/s, which is enough for good small and mid-sized models but not for the largest open-weight ones. Our explainer on choosing a small model for your laptop has the memory math this piece builds on.

This explainer is based on Apple's October 27, 2026 newsroom releases and the MacBook Pro and iMac product and tech-specs pages, Apple's October 28, 2026 release for the MacBook Pro with PROLED XDR Display, and Apple's November 19, 2025 machine-learning research post on MLX and the M5, checked October 29, 2026. Apple's AI performance figures come from Apple's own testing and are labeled vendor-reported.

The short version

• What launched: the 14-inch MacBook Pro with M6, from $1,999, and the iMac with M6, from $1,499. Pre-orders opened October 28; both arrive November 6, 2026.
• AI hardware: a "Dual 16-core Neural Engine," which Apple says is up to 1.9x faster than M5's, and a Neural Accelerator in each of the GPU's cores.
• Memory: 16GB standard, up to 32GB on the MacBook Pro and the 12-core iMac, 24GB on the 10-core iMac.
• Bandwidth: 153GB/s with 16GB and 170GB/s with 24GB or 32GB, according to the MacBook Pro tech specs, which shapes how fast a model writes its answer.
• What fits: 4-bit models up to about 14B parameters at 16GB, and 20B to 30B mixture-of-experts models at 24–32GB. 70B-class models don't fit.

What did Apple announce on October 27, 2026?

Two Macs with the base M6 chip, and no M6 Pro or M6 Max. Apple's newsroom lists the 14-inch MacBook Pro with M6 and the iMac with M6 as its October 27 releases.

Last verified: October 29, 2026

Storage starts at 1TB on the MacBook Pro and 256GB on the iMac, with up to 4TB and 2TB respectively. The 14-inch MacBook Pro is still listed with M5 Pro and M5 Max chips. On Apple's tech-specs page, the M5 Pro offers up to 64GB and 307GB/s, and the M5 Max up to 128GB and 614GB/s. On October 28, Apple also introduced a MacBook Pro with PROLED XDR Display, a tandem OLED screen, in 14- and 16-inch sizes with M5 Pro or M5 Max chips and up to 128GB of unified memory, from $2,899. That launch had no M6 Pro or M6 Max either, and Apple's release doesn't say whether the new models replace the earlier M5 Pro and M5 Max MacBook Pros.

What does Apple claim for AI performance?

All of these figures are vendor-reported, from Apple's own tests:

• MacBook Pro versus M5: "up to 25 percent faster performance for on-device AI workloads, like LLM prompt processing," and a Neural Engine that "delivers up to 1.9x faster performance than M5." The comparison machine was a 14-inch MacBook Pro with M5, a 10-core CPU and GPU, 32GB of memory and a 4TB SSD.
• MacBook Pro versus M1: "up to 13x faster time to first token performance for LLM prompt processing" than the 13-inch MacBook Pro with M1, tested in September 2026.
• iMac versus M4 and M1: "up to 5x faster on-device AI performance than M4" and up to 12x faster than M1 for prompt processing in LM Studio. The iMac product page's footnote says that test used an 8K-token prompt and a 14-billion-parameter model at 4-bit quantization. (There was no M5 iMac, so the comparison skips a generation.)

These claims measure time to first token: how long the Mac takes to read your prompt before it starts answering. They don't measure how fast it then writes the answer, and Apple gives no figure for that.

Why does memory matter more than the Neural Engine?

Because the model has to fit, and because writing the answer is limited by memory speed. Apple's own machine-learning researchers said so in their November 19, 2025 post on running LLMs with its MLX framework on the M5: "In LLM inference, generating the first token is compute-bound," while "generating subsequent tokens is bounded by memory bandwidth, rather than by compute ability."

That splits the M6's improvements in two. Apple's AI claims are about the compute-bound part, reading long prompts and documents, where the GPU's Neural Accelerators help. The speed at which text appears depends mostly on bandwidth. The M6 with 24GB or 32GB has 170GB/s; Apple's MLX post gave the M5 153GB/s and the M4 120GB/s. On the M5, Apple measured prompt processing up to about 4x faster than the M4 but generation only 19% to 27% faster, roughly in line with the bandwidth gain.

Unified memory is why Macs do well here. The CPU and GPU share one pool, so a 32GB Mac can give a model far more memory than a PC with a 12GB graphics card, as our laptop model guide explains. Whether a given app uses the Neural Engine at all is another matter: Apple's post describes MLX using the Neural Accelerators inside the GPU, and our explainer on what an NPU actually does covers why a chip's AI-engine headline number says little about local models.

What size AI model can an M6 Mac run?

Start from Apple's own measurements. In its MLX tests on a 24GB M5 MacBook Pro, with a 4,096-token prompt, Apple reported these memory figures:

Last verified: October 28, 2026

Quantization makes the biggest difference: Qwen3 8B used 17.46GB at 16-bit and 5.61GB at 4-bit. Apple wrote that "the MacBook Pro 24GB can easily hold a 8B in BF16 precision or a 30B MoE 4-bit quantized." Those figures cover a 4K-token prompt; longer context needs more memory for the attention cache, which M7 shows can exceed the model itself at long lengths. Your operating system and apps need room too. M7's rule of thumb is to plan on about two-thirds of RAM for the model and its context on a machine you're also working on.

Applied to the M6 options, as a rough guide rather than a measurement:

• 16GB: two-thirds is about 10.7GB, enough for 4-bit models up to about 14B parameters at modest context.
• 24GB: about 16GB, enough for gpt-oss-20b or a 30B mixture-of-experts model at short context.
• 32GB: about 21GB, the same models with more room for context.

For a 70B model at 4-bit, M7's formula (parameters × bits per weight ÷ 8, using 4.5 bits for common 4-bit formats) gives about 39GB for the weights alone. That's beyond every M6 configuration. It's where the M5 Max MacBook Pro, with up to 128GB, or a dedicated box such as Lenovo's AI Host P7 comes in.

How fast will local models run on an M6?

Apple publishes no tokens-per-second figure for the M6, and speeds depend on the model, quantization, context length and software. Bandwidth sets a rough ceiling: each new token requires reading roughly the whole model from memory. On a 24GB or 32GB M6, dividing 170GB/s by the 9.16GB Apple measured for a 4-bit 14B model gives an upper bound of about 19 tokens per second. Real speeds land below such ceilings. Treat this as arithmetic, not a measurement; tools like Ollama report actual speeds on your own machine, as our Ollama guide describes.

The practical upshot: the M6 should feel noticeably faster than older Macs when it reads long prompts, which is what Apple's claims measure. Answer-writing speed improves less, in line with the modest bandwidth gain over M5.

Should you buy an M6 MacBook Pro or iMac for local AI?

It depends on the models you want to run:

• Small and mid-sized models (up to about 14B at 4-bit): a 16GB M6 handles them. 24GB gives room for longer context.
• 20B to 30B mixture-of-experts models: choose 24GB or 32GB.
• Larger models, 70B and up: the M6 isn't built for them. The M5 Pro (up to 64GB) and M5 Max (up to 128GB) MacBook Pros on Apple's tech-specs page have more memory and more bandwidth, and Apple's October 28 PROLED XDR MacBook Pro uses the same two chips.

What's confirmed, reported and uncertain

• Confirmed by Apple: the models, prices, dates, core counts, memory options, bandwidth figures and the "Dual 16-core Neural Engine" name, and the October 28 PROLED XDR MacBook Pro's chips, 128GB ceiling and $2,899 starting price.
• Vendor-reported: every AI performance multiple (1.9x, 25%, 13x, 12x, 5x) and the MLX memory and speedup figures.
• Illustrative calculations: the two-thirds memory budgets, the 70B weight estimate and the tokens-per-second ceiling.
• Uncertain: independent tokens-per-second results for M6 Macs, what the "Dual" Neural Engine design means in practice, memory upgrade prices, which Apple's pages don't list, and whether the PROLED XDR models replace the earlier M5 Pro and M5 Max MacBook Pros.