New MacBook Pro for Local AI: How Much RAM Do You Need?

The short version

Apple updated the MacBook Pro twice this week. On 27 October it announced a 14-inch model with the new M6 chip, and on 28 October the long-rumoured touchscreen MacBook Pro with an OLED “PROLED XDR” display. If you want a laptop that runs AI models on the machine itself, rather than in the cloud, the headline features matter less than one number: memory.

• The touchscreen doesn’t change what you can run. The OLED models use the M5 Pro and M5 Max chips Apple launched in March, not the M6.
• Skip the 16GB M6 if AI is the reason you’re buying. Only the 24GB and 32GB versions get the faster 170GB/s memory; 16GB stays at 153GB/s.
• Memory decides model size; bandwidth decides speed. More of both costs money, so decide what size of model you actually want to run first.

What Apple announced

14-inch MacBook Pro with M6. A 12-core CPU, a 12-core GPU with a Neural Accelerator in each core, and 16GB of unified memory as standard, configurable to 32GB. It starts at $1,999 with 1TB of storage, ships on 6 November, and keeps the mini-LED Liquid Retina XDR display. Apple says it is up to 25% faster than M5 for on-device AI work such as processing an LLM prompt, and up to 1.3 times faster to produce the first token.

Touchscreen MacBook Pro. 14-inch and 16-inch models with a tandem OLED panel Apple calls PROLED XDR, touch input, and up to 128GB of memory and 16TB of storage at the top end. MacRumors reports the OLED versions cost about $400 more than comparable mini-LED models, and they use the M5 Pro and M5 Max.

The M5 Pro and M5 Max, for reference. When Apple launched them in March, the M5 Pro went up to 64GB of memory at 307GB/s and the M5 Max up to 128GB at 614GB/s.

Why memory is the number that matters

A model has to fit in memory before it can run at all. Our guide on how much memory an AI model needs has the rule of thumb: at the common 4-bit setting, a model takes roughly 0.6 bytes per parameter, plus room for the conversation (the KV cache) and a gigabyte or two for the software around it. So an 8-billion-parameter model needs about 5GB, a 32B model about 19GB and a 70B model about 42GB, before context.

On a Mac, there’s a second limit. By default macOS only lets the GPU use part of the unified memory, commonly reported as somewhere between two-thirds and three-quarters, and macOS and your open apps need their share too. A 16GB Mac does not give a model 16GB.

Then there is speed. When a model writes its answer, it reads through its weights for every token, so memory bandwidth sets a ceiling. Divide bandwidth by model size for a rough upper limit: 170GB/s over a 5GB model is at most around 34 tokens a second, and over a 19GB model around 9. Real speeds come in lower. Apple’s M6 claims are mostly about the first token, which depends on raw compute; how fast the rest of the answer arrives depends on bandwidth.

What each configuration can realistically run

These are our estimates from the rule above, for 4-bit models with a modest context. They are calculations, not benchmarks; we haven’t tested the new machines.

“Large MoE models” means mixture-of-experts models such as Meta’s Llama 4 Scout, which our earlier guide puts at roughly 60 to 70GB at 4-bit. They answer faster than their size suggests, because only part of the model works on each token. Some newer tools go further and stream parts of these models from the SSD, which we covered in running an AI model bigger than your RAM, but that is still experimental.

Which one should you buy?

• You use ChatGPT or Claude in a browser. None of this matters. Buy for the screen, battery and price.
• You want to try small local models. The M6 with 24GB is the sensible floor. It gets the faster memory and room for 8B to 14B models alongside your other apps.
• You want models that are useful for real work, such as coding. Look at the M5 Pro with 64GB. More memory and nearly double the bandwidth of the M6 matter more than the newer chip name.
• You want 70B-class or large MoE models on a laptop. That is M5 Max territory, and 128GB is the only configuration with comfortable headroom.
• You don’t need to carry it. A desktop gets you the same chips and memory for less. Our roundup of local AI hardware this autumn compares the Mac Studio with NVIDIA’s RTX Spark machines.

Unified memory can’t be upgraded later, so if local AI is part of the plan, it is the one place not to save money.

What we don’t know yet

• Independent speed tests. Apple’s figures come from its own testing, and reviews haven’t arrived yet.
• How the touchscreen models’ thinner design affects sustained performance. Long AI jobs keep the chip busy, so cooling matters.
• Full official pricing for the OLED models in every configuration. We’ve relied on reported figures here.
• Whether the M6 Pro and M6 Max arrive soon, and with more memory. Apple hasn’t said.

Sources

• Apple Newsroom: the new 14-inch MacBook Pro with M6
• MacRumors: M6 MacBook Pro’s 170GB/s bandwidth skips the 16GB model
• MacRumors: everything Apple announced this week
• Tom’s Guide: touchscreen MacBook Pro event coverage
• MacRumors: MacBook Pro with M5 Pro and M5 Max (March 2026)

For disclosure: we haven’t tested these machines, and the model-size figures are our own estimates from a general rule. Prices are US list prices at announcement and may differ in your country. We use AI tools to help research and draft, and we check facts and sources before publishing.

More from Offgrid Studio

How Much Memory Does an AI Model Need?
Local AI Hardware in Autumn 2026: What You Can Actually Run at Home
Nemotron 3.5 ASR vs Whisper: Which Speech-to-Text to Use