Home / Guides / Troubleshooting / Apple
macOS
Written by Jakub Rusinowski · Last updated September 1, 2026
Founder, LLM Configurator — AI educator & workshop leader on local LLM deployment
LM Studio system requirements: Apple Silicon Mac (M1/M2/M3/M4), macOS 14.0 or newer.
Intel Macs are not supported.| If you see | The cause is | Go to |
|---|---|---|
| The LM Studio installer refuses, or the download page offers no Intel build | LM Studio requires Apple Silicon and macOS 14+ | Fix 1: confirm which Mac you have, then Fix 2 |
uname -m prints x86_64 | This is an Intel Mac — no Apple Silicon path applies | Fix 2: use tools that still support it |
| Ollama runs but every model is slow | You are on CPU inference; there is no fast path here | Fix 3: size the model for CPU speeds |
| You have an AMD dGPU in a 16-inch and expect it to help | Metal on those parts lacks the memory bandwidth this workload needs | Fix 4: expect CPU-class results |
| You need real speed on a large model | The machine cannot provide it at any setting | Fix 6: rent GPU time, or move to Apple Silicon |
Worth putting the answer in the first sixty words: LM Studio on macOS requires Apple Silicon and macOS 14.0 or newer, and Intel Macs are not supported. No setting changes that. What follows is what does still work on an Intel Mac, and how fast to expect it to be.
Local inference is bound by memory bandwidth, and that is exactly where Apple Silicon differs from an Intel Mac. On an M-series chip, CPU and GPU share one high-bandwidth memory pool, so weights are read fast and never copied across a bus — which is why a MacBook Air can outrun a far more expensive Intel machine at this one task. An Intel Mac has conventional system memory feeding the CPU, and either Intel integrated graphics or an AMD discrete GPU with its own separate, comparatively small memory. Metal exists on those GPUs, but neither gives you the bandwidth story that makes Apple Silicon good here. An Intel Mac is a CPU-inference machine, and the useful question is not how to unlock its GPU but which models are worth running on its CPU.
Before anything else, settle the architecture question — Rosetta 2 makes this less obvious than it should be, because an Intel-built terminal on an M-series Mac reports x86_64 too. uname -m in a native terminal, plus the chip line from system_profiler, gives you the truth.
# zsh
uname -m # arm64 = Apple Silicon, x86_64 = Intel
sysctl -n machdep.cpu.brand_string
system_profiler SPHardwareDataType | grep -E "Model Name|Chip|Processor|Memory"
sw_vers # your macOS versionProcessor line naming an Intel Core chip confirms it. If you see a Chip line naming an M-series part, this is the wrong page — the MLX and swapping guides are the relevant ones.Ollama still runs on Intel Macs, on the CPU, and remains the simplest route — it handles the model store and the chat loop without you building anything. llama.cpp builds and runs from source, and gives you more control over threads and context. Metal is enabled by default in llama.cpp on macOS; on an Intel Mac you may want it off, since the CPU path is often the faster of the two here — -DGGML_METAL=OFF at configure time does that. Benchmark both before settling.
# zsh — Ollama, CPU inference
brew install ollama
ollama run llama3.2:3b
# llama.cpp from source, CPU-only build
git clone https://github.com/ggml-org/llama.cpp && cd llama.cpp
cmake -B build -DGGML_METAL=OFF
cmake --build build --config Release -j
./build/bin/llama-cli -m model.gguf -p "hello" -t 8ollama run llama3.2:3b "Write one sentence about the sea." --verboseOn an Intel Mac the constraint is throughput rather than capacity, which inverts the usual advice. 3B to 8B at Q4 is the realistic band, and single-digit tokens per second is the realistic speed. A 2019 16-inch with 32 GB will run an 8B usably for chat — slower than you would like, fast enough to be useful — and will not run a 70B in any meaningful sense, however much memory it has. Start at 3B, confirm the speed is tolerable, and move up only if it is.
# zsh — start small and measure
ollama run llama3.2:3b --verbose # note eval rate in tokens/s
ollama run llama3.1:8b --verbose # compare
# llama.cpp: match threads to physical cores, not logical
sysctl -n hw.physicalcpu
./build/bin/llama-cli -m model.gguf -t $(sysctl -n hw.physicalcpu)The 15- and 16-inch Intel MacBook Pros shipped with AMD discrete graphics, and it is reasonable to assume that helps. In practice those parts have a few gigabytes of dedicated memory and bandwidth well below what a modern inference GPU provides, and the ROCm ecosystem does not cover macOS at all. Metal can reach them, but the result lands in the same performance neighbourhood as the CPU. Test it if you like — just calibrate expectations before you spend an evening on it.
# zsh — what graphics does this machine actually have?
system_profiler SPDisplaysDataType | grep -E "Chipset|VRAM|Vendor"CPU inference makes prompt processing painfully visible: a long document is re-read at CPU speed before a single token comes back. Cutting the context length reduces both the memory footprint and the per-request wait, and on this class of machine that trade is almost always worth taking. Ollama's default is 4096; there is rarely a reason to raise it here.
# zsh — keep the window modest on CPU
OLLAMA_CONTEXT_LENGTH=2048 ollama run llama3.2:3bThere is no configuration that makes an Intel Mac fast at this, so the honest choice is between three things. Accept CPU speeds with a 3–8B model, which is genuinely fine for chat, drafting and light coding help. Rent GPU time by the hour, which is the right answer for occasional heavy work — a few hours a month on a rented card costs less than a machine and gives you hardware you could not buy at that price. Or move to Apple Silicon, where even an entry-level M-series machine transforms this workload. Cost is the thing to work out before deciding, not after.
If you are weighing a new Mac, the buying guide covers which memory tier matters for the model sizes you want. If you are leaning toward renting, the cost calculator compares hourly GPU rental against buying outright for your actual usage.
Include this when you report it
uname -m, sysctl -n machdep.cpu.brand_string and sw_verssystem_profiler SPHardwareDataType memory and model linesollama run --verbose on a 3B modelNo. LM Studio on macOS requires Apple Silicon (M1/M2/M3/M4) and macOS 14.0 or newer, and Intel Macs are not supported by current builds. Ollama and llama.cpp still run on Intel, on the CPU.
3B to 8B at Q4, at single-digit tokens per second. A 2019 16-inch with 32 GB runs an 8B usably for chat and will not run a 70B in any meaningful sense, no matter how much memory it has — the limit is bandwidth, not capacity.
Not much. Those parts have a few gigabytes of dedicated memory and bandwidth well below a modern inference GPU, and ROCm does not cover macOS. Metal can use them, but results land close to CPU performance. Benchmark it if curious; do not plan around it.
Try both. Metal is enabled by default on macOS; on an Intel Mac the CPU path is often as fast or faster, and -DGGML_METAL=OFF at configure time gives you a CPU-only build to compare against. Ten minutes of benchmarking settles it for your machine.
If local inference is a regular part of your work, the difference is large — unified memory bandwidth is exactly what this workload needs. For occasional heavy jobs, renting GPU time by the hour is usually cheaper than a new machine. Work out your actual hours first.