On 27 October Apple announced a 14-inch MacBook Pro with its new M6 chip, and much of the pitch was about AI. Apple says the machine is built for "running LLMs locally, all while preserving user privacy", and claims "up to 25 percent faster performance for on-device AI workloads, like LLM prompt processing, compared to M5". "macbook pro" has been trending on Google since.
This is not a review, and nothing here is a recommendation to buy one. The more useful question for a business is the one the M6 MacBook Pro AI pitch raises: can you run AI models on staff laptops instead of sending work to a cloud provider, and when is that worth doing? The answer depends less on the chip than on memory, on the task and on who maintains the result.
What Apple actually announced
The figures below are Apple's, from its own announcement, and every performance number is a vendor claim. Apple's footnotes say the testing was its own, run in September and October 2026, and that the M5 comparisons were against a 14-inch M5 MacBook Pro with 32GB of memory.
The chip. A 12-core CPU and a 12-core GPU, two more cores of each than before, with "Neural Accelerators in each core" of the GPU, plus what Apple calls a "Dual 16-core Neural Engine".
The AI claims. Up to 25 percent faster AI workloads than M5, and "up to 13x faster time to first token performance for LLM prompt processing" compared with a 13-inch M1 MacBook Pro. Apple also says the Neural Engine is up to 1.9 times faster than M5's.
Memory. "16GB of unified memory standard, configurable up to 32GB", with bandwidth of up to 170GB/s. TechCrunch puts the M5's figure at 153GB/s.
Price and dates. From $1,999 in the US, or $1,899 with education pricing. Pre-orders opened on 28 October and the machine is available from Friday 6 November.
John Gruber at Daring Fireball lists the memory options at $200 extra for 24GB and $400 for 32GB, and notes that prices "seem unchanged from the post-June-price-hike M5 models". That hike is part of the story. In June Apple raised the entry 14-inch MacBook Pro from $1,699 to $1,999, and a spokesperson told Bloomberg that "the rapid expansion of AI data centers has created an extraordinary surge in demand for memory and storage." The memory that local AI needs most is exactly the part that has become expensive.
What on-device AI means, and what the "25% faster" covers
Running a model locally means its weights sit on the laptop and the prompt, the document and the answer never leave the machine. Tools such as Ollama and Apple's own MLX framework make this a download and a command rather than a research project. Ollama's documentation says plainly: "Ollama runs locally. We don't see your prompts or data when you run locally."
Apple's headline claim deserves a closer reading. A model answering a question does two different jobs. First it reads the whole prompt, which decides how long you wait before the first word appears. Then it writes the answer one token at a time. Databricks' engineering guide explains that generating the first token "is typically compute-bound, while subsequent decoding is memory-bound", and that generation at small batch sizes "is a GPU memory bandwidth-bound problem".
Apple's 25 percent and 13x figures are about the first job, which matters most for long documents. The speed of writing the answer is governed mainly by memory bandwidth, which by the figures above rose about 11 percent (170GB/s against 153GB/s). Expect a machine that starts answering sooner and writes somewhat faster, not one that is a quarter faster at everything.
For scale, community benchmarks collected by the llama.cpp project record a base M5 writing about 32 tokens a second with a 7-billion-parameter model at 8-bit precision, faster than anyone reads. There are no independent M6 results yet.
How big a model fits in 16, 24 or 32GB
Memory, not the processor, decides which models a laptop can run at all. A model's weights must fit in memory, and the size depends on how many parameters it has and how many bits each is stored in.
The same llama.cpp benchmark gives a clean example. One 7-billion-parameter model takes 12.55 GiB at 16-bit, 6.67 GiB at 8-bit and 3.56 GiB at 4-bit. That gives a rough rule of thumb: about 2 bytes per parameter at 16-bit, about 1 byte at 8-bit and a little over half a byte at 4-bit, the precision Epoch AI assumes in the estimate below. Treat it as approximate, because three other things need memory too.
The conversation itself. The model keeps a working cache of the text it is handling, and Ollama's documentation notes that the memory required grows with the context length multiplied by the number of requests handled in parallel. In other words, longer documents and more simultaneous requests cost more.
Everything else that is open. Unified memory is shared with email, the browser and calls.
A ceiling on GPU memory. macOS sets a system limit on how much memory the GPU keeps resident. Apple's MLX documentation shows how to raise it with a system setting, and says the limit "should remain strictly less than the total memory size". That is not something to push across a fleet of laptops without thought.
Epoch AI, working the same arithmetic for graphics cards, estimated that a 32GB card can hold models of about 40 billion parameters at 4-bit, assuming an 8,000-token context and some overhead. A 32GB laptop has less to spare, because it is also running everything else. With those caveats, and using download sizes from Ollama's library, the picture looks roughly like this.
| Memory | Upgrade price (US) | Realistic local models | Example sizes |
|---|---|---|---|
| 16GB | included | Small models, up to about 8–12B at 4-bit | Qwen3 8B, 5.2GB; Gemma 3 12B, 8.2GB |
| 24GB | +$200 | Mid-sized models around 14B with room for longer documents | Qwen3 14B, 9.3GB |
| 32GB | +$400 | Models in the 27–32B range, with little room left over | Gemma 3 27B, 17GB; Qwen3 32B, 20GB |
Sources: model sizes from Ollama's Qwen3 and Gemma 3 pages; upgrade prices from Daring Fireball.
Some models are now designed for this budget. OpenAI's open-weight gpt-oss-20b, released under an Apache 2.0 licence, is described on its model card as able to "run within 16GB of memory" and aimed at "local or specialized use cases". On a 16GB laptop that leaves nothing spare.
Local AI for business: the trade-offs against cloud models
Privacy and data residency. The strongest argument. If a contract or a personnel file never leaves the laptop, there is no third-party processor to assess and no provider retention policy to rely on. For data that may not leave the building, that can decide whether AI is usable at all.
Cost per seat against cost per token. A cloud model charges for use. A local model costs mostly up front: the extra memory and the time to set it up. If a team summarises documents all day, a fixed $400 upgrade per laptop may beat a metered bill. If they use AI twice a week, it will not. Compare the cost of a finished task, not the price of a request.
Quality. This is where the coverage is most optimistic. Epoch AI finds that the best open-weight models trail the frontier by an average of about four months, but those leading open models are generally far too large for a laptop. The more relevant figure is for models that fit on one consumer graphics card: they match frontier capability with an average lag of 6 to 12 months, and Epoch cautions that small models may be tuned to benchmarks, so the real gap may be longer. A laptop model is roughly where the best cloud models were a year ago: enough for many tasks, not for hard reasoning.
Maintenance. A cloud model is someone else's job to update. Local models mean choosing, pinning, updating and re-testing a model on every machine, indefinitely.
Security on the endpoint. Local AI removes one risk and adds others. There is no cloud key to steal, which matters given how often API keys are leaked and abused. But a model file is software from the internet. Hugging Face warns that loading pickle-format model files can allow "dangerous arbitrary code execution attacks" and advises loading models only from sources you trust. And a local model server is a network service: Ollama "binds 127.0.0.1 port 11434 by default", but one changed setting exposes it to the network. Both belong in your device policy.
Which tasks suit a model running locally on a Mac
Good fits. Drafting and rewriting internal text. Summarising and extracting from sensitive documents, where privacy is the reason for doing it locally at all. Searching or questioning a folder of the user's own files. Coding help when offline, or in codebases that must not be sent to a third party. All of these tolerate an occasional weaker answer because a person reads the output.
Poor fits. Anything that needs the strongest reasoning available, such as complex analysis, legal or financial judgement, or difficult code. Anything shared by a team or connected to a business system, which belongs on a server you manage rather than on one person's laptop. Very long documents on a 16GB machine. And anything where nobody will check the output, which is a problem whether the model is local or not.
Often the sensible design is both, with a clear rule about which data may go where.
How to pilot local AI before buying anything
Start with the task, not the hardware. Pick one job where privacy or cost is the reason to go local, and write down what a good result looks like. If you cannot name the task, the question of whether you need AI for it at all comes first.
Test on machines you already have. An Apple-silicon Mac with 16GB can run the small models in the table. Put 20 or 30 real examples through two or three of them and compare the results with the cloud model you would otherwise use.
Measure what matters. How often the answer is usable unedited, how long a typical document takes, and how much memory is left for everything else.
Decide the rules before rollout. Which models are approved, where they are downloaded from, who updates them and whether the local server may listen on the network.
Only then is it worth asking whether 24GB or 32GB earns its upgrade price.
The short version
The M6 MacBook Pro makes local AI faster, by Apple's own claims mostly in how quickly a model reads its input. It does not change the constraint: memory decides which models fit, 16 to 32GB holds small and mid-sized ones, and they trail the best cloud systems. For private, routine, human-checked work, a laptop can now do a useful share of the job. The decision rests on the task and the data, not on the chip.