Mac mini for local LLMs in 2026: which config actually runs what
Which Mac mini runs local LLMs well in 2026. The two specs that decide it, what 16GB, 24GB and 48GB each fit, and the honest ceiling on 70B models.
The Mac mini is the cheapest way into serious local language model work, which is why the question keeps coming up: which one do you actually buy? The answer comes down to two numbers on the spec sheet, and neither of them is the one most buying guides lead with. Memory bandwidth sets how fast a model generates text. Unified memory size sets which models you can load at all. Everything else, including GPU core count, matters far less than either.
We build recal, a local-first assistant that runs models on the user's own Mac, so these constraints are the ones we design around rather than ones we read about. Nothing below is a recommendation to buy recal, and the hardware advice stands on its own.
Key takeaways
- The M4 moves 120GB/s. The M4 Pro moves 273GB/s. Since token generation is memory-bandwidth bound, that gap is roughly the speed difference you will feel, and it is much larger than the difference in raw compute.
- Memory size decides what loads; bandwidth decides how fast it runs. You need to clear both bars, and they are set by different chips.
- The Mac mini tops out at 48GB of unified memory, on the M4 Pro. There is no 32GB M4 Mac mini, despite what several buying guides currently claim.
- macOS does not give the GPU all of your RAM. Budget roughly two thirds to three quarters of it, which is what turns a 16GB machine into an 8B machine.
- A 70B model at 4-bit does not comfortably fit in 48GB. Guides that say it does are quoting weight size and forgetting the context cache.
What are the current Mac mini configurations?
As of August 2026 the Mac mini still ships with the M4 and M4 Pro. The M5 arrived in other Macs and the mini has not been refreshed yet, so anyone waiting for an M5 mini is still waiting. Apple's own specifications list the lineup like this:
| M4 | M4 Pro | |
|---|---|---|
| GPU cores | 10 | 16, configurable to 20 |
| Memory bandwidth | 120GB/s | 273GB/s |
| Unified memory | 16GB, configurable to 24GB (one model ships at 24GB) | 24GB, configurable to 48GB |
| Practical model ceiling | 8B comfortably, 14B on 24GB | 32B comfortably, 70B only with compromises |
The single most important line there is memory bandwidth. The M4 Pro moves weights more than twice as fast as the M4, and that is the number that governs how quickly text appears on screen.
Why does memory bandwidth matter more than the chip name?
Generating a token requires reading the model's weights out of memory. Every token, all of them. So for a single-user local setup, decode speed is limited by how fast memory can be read, not by how fast the processor can multiply. This is why a chip with modest compute and fast memory beats a chip with strong compute and slow memory for this specific job.
You can estimate the ceiling with division. Take the memory bandwidth and divide by the size of the model in memory:
- An 8B model at 4-bit occupies roughly 4.7GB. On the M4 that is 120 divided by 4.7, about 25 tokens per second as an absolute ceiling. On the M4 Pro it is about 58.
- A 32B model at 4-bit occupies roughly 18GB. On the M4 Pro that is about 15 tokens per second at the ceiling.
- A 70B model at 4-bit occupies roughly 40GB, giving about 7 tokens per second on the M4 Pro before any real-world losses.
Real throughput lands lower, usually somewhere around 60 to 70 percent of that ceiling once you account for the attention cache, sampling, and the fact that nothing runs at theoretical peak. Treat these as upper bounds you cannot exceed rather than benchmarks you should expect. The useful part is the ratio: it tells you what changing the model size or the chip will do before you spend anything.
How much RAM do you need for a local LLM on a Mac mini?
More than the model file, and that is where most estimates go wrong. Three things compete for unified memory: the model weights, the attention cache that grows with your context length, and macOS itself along with everything else you have open.
macOS also does not hand the GPU your full RAM. The system reports a recommended maximum working set that lands around two thirds of total memory on smaller configurations and closer to three quarters on larger ones. You can raise it with the iogpu.wired_limit_mb tunable, but you are then gambling that nothing else needs the memory, and the failure mode is the system swapping or the process dying rather than a graceful slowdown.
Working backwards from that:
- 16GB gives the GPU roughly 10 to 11GB. That is an 8B model at 4-bit with room for a real context window, or a 14B if you quantize hard and keep context short. It is a genuine local-LLM machine, just not a large-model one.
- 24GB gives you roughly 16 to 18GB. This is where 14B models become comfortable and a 32B becomes possible at low quantization with modest context. For most people this is the sweet spot on price.
- 48GB gives you roughly 34 to 36GB. A 32B model runs comfortably with a long context. A 70B at 4-bit needs about 40GB for weights alone, so it does not fit inside the default limit at all.
Is a 70B model realistic on a Mac mini?
Not really, and this is the claim worth being blunt about. A 70B model at 4-bit is roughly 40GB of weights. The largest Mac mini has 48GB total, of which macOS will offer the GPU around 36GB by default. So you are already short before the context cache allocates a single byte.
You can make it run. Raise the wired memory limit, quantize to 3-bit or lower, and keep context small. What you get back is a model that has lost measurable quality to the quantization, generating at roughly 5 tokens per second, on a machine with nothing left over. If running 70B locally is your actual requirement, a Mac Studio with more unified memory is the honest answer, and a Mac mini is the wrong purchase.
For the record, this is where several current buying guides go astray. At least one widely circulated 2026 recommendation suggests a 32GB M4 Mac mini as the value pick. That configuration does not exist. The M4 mini stops at 24GB, and 32GB is not an option at any price.
Do GPU cores matter for running local models?
Less than you would think for generating text, and more than you would think for reading it.
Token generation is bandwidth bound, so extra GPU cores mostly sit idle waiting on memory. Upgrading the M4 Pro from 16 to 20 GPU cores will not make your chatbot noticeably faster to reply.
Prompt processing is different. When the model reads a long document before answering, it processes tokens in parallel, and that stage is compute bound. If you paste large files, work with long transcripts, or run retrieval over a big context, extra GPU cores genuinely help there. So the core-count upgrade is worth it for document-heavy work and close to pointless for conversation.
Which Mac mini should you buy for local LLMs?
- Buy the 16GB M4 if you want to learn what local models can do, run an 8B assistant, and keep the cost down. It works, and it will teach you more about the constraints than any article will.
- Buy the 24GB M4 if you want the best value. It opens up 14B models, which is the tier where local output starts feeling genuinely useful rather than merely impressive, and it costs far less than stepping up to the Pro.
- Buy the 48GB M4 Pro if local models are the point of the machine. You get both the bandwidth to run them quickly and the memory to run 32B comfortably. This is the configuration to get if you are choosing once and keeping it.
- Do not buy the 24GB M4 Pro for this purpose specifically. You pay for the bandwidth and then cannot load the larger models it would run well. Either save money with the M4 or spend up to 48GB.
If you want to know which runtime to put on it once it arrives, our MLX vs llama.cpp comparison covers the engine choice, and Ollama vs LM Studio covers the app you will actually click on. For picking the model itself, see our best local LLM apps for Mac.
Where recal fits
recal runs on the same hardware and lives with the same limits. It is a local-first assistant for the Mac that watches how you work, on-device, and starts doing the busywork itself, surfacing finished work for you to approve or reject. The reason the numbers above matter to us is that a private assistant is only private if the model stays on your machine, which means it has to fit in your unified memory and run at a speed you will tolerate. There is no server to fall back on, by design, so the memory budget is the product constraint.
That is also why we care about the bandwidth number more than the marketing. An assistant that thinks at 5 tokens per second is not one you will let near your actual work.
Frequently asked questions
Is the Mac mini good for local LLMs? Yes, for models up to about 32B. Unified memory means the GPU can address the whole memory pool, which is why a mini competes with discrete graphics cards that have far less video memory. Its limit is total memory size, not architecture.
Can a 16GB Mac mini run a local LLM? Yes. An 8B model at 4-bit fits comfortably with room for context, which covers general assistance, summarising, and code completion. You will not run 32B or larger.
Is the M4 Pro worth it over the M4 for AI? For local models, yes, because 273GB/s against 120GB/s is more than double the decode ceiling. For most other work the gap is much smaller, so this is one of the few cases where the Pro upgrade is justified by a single number.
Does a Mac mini beat a gaming PC for local LLMs? It depends on model size. A discrete card with 12 or 16GB of video memory will beat the mini on anything that fits in that memory. Once the model exceeds the card's memory, the mini's larger unified pool wins, because spilling to system memory over PCIe is dramatically slower.
Should I wait for the M5 Mac mini? If you can. A refresh has been reported as likely before the end of 2026, and a bandwidth increase would matter more for this workload than a compute increase. If you need the machine now, the M4 Pro is not going to stop being capable.
Written with AI assistance and reviewed by the recal team. Specifications are from Apple's published Mac mini technical specifications as of August 2026. Throughput figures are calculated bandwidth ceilings, not measured benchmarks, and are labelled as such above.