What works on an Intel Mac in 2026

macOS 14 Sonoma supports Intel MacBook Pros and MacBook Airs from 2018 on, the 2018 Mac mini, iMacs from 2019, the iMac Pro and the 2019 Mac Pro (Apple’s compatibility list). Apple has also said that macOS Tahoe 26 is the last macOS release for Intel Macs (Six Colors, June 2025), so app support will thin out from here.

i7/i9 tokens per second, by Mac

On a CPU, as on a GPU, generating each token means reading the model’s weights from memory, so memory bandwidth sets a ceiling. The table applies the LLMCheck index’s bandwidth model to each Mac’s memory as Apple specifies it, for 4-bit (Q4_K_M) models. An i7 and an i9 with the same memory share the same ceiling; the i9’s extra cores help it get closer to it.

Mac (16 GB) and memory3.8B (Phi-4 Mini)8B (Llama 3.1 8B)14B (Phi-4)
MacBook Pro 16-inch (2019), i7 or i9
DDR4-2666, 42.7 GB/s
estimatedup to ~15 tok/sestimatedup to ~7 tok/sestimatedup to ~4 tok/s
MacBook Pro 15-inch (2018–2019), i7 or i9
DDR4-2400, 38.4 GB/s
estimatedup to ~14 tok/sestimatedup to ~7 tok/sestimatedup to ~4 tok/s
MacBook Pro 13-inch (2020, four ports), i5 or i7
LPDDR4X-3733, 59.7 GB/s
estimatedup to ~21 tok/sestimatedup to ~10 tok/sestimatedup to ~6 tok/s
iMac 27-inch (2020), i5, i7 or i9
DDR4-2666, 42.7 GB/s
estimatedup to ~15 tok/sestimatedup to ~7 tok/sestimatedup to ~4 tok/s
For comparison: MacBook Air M1
LPDDR4X, unified, 68 GB/s
estimated~24 tok/sestimated~11 tok/sestimated~7 tok/s
For comparison: Mac mini M6 16 GB
unified, 153 GB/s
estimated~51 tok/sestimated~25 tok/sestimated~14 tok/s

Memory specifications: Apple’s technical specifications for the MacBook Pro 16-inch (2019), MacBook Pro 13-inch (2020) and iMac 27-inch (2020), and EveryMac for the 15-inch. Bandwidth is the theoretical figure for a 128-bit memory bus.

Read the Intel rows as ceilings. The index covers Apple Silicon only, and its bandwidth model was fitted to Apple GPUs. Applied to an Intel Mac it gives an upper estimate: a CPU, especially a four-core one or a hot laptop, usually decodes below it. Comfortable reading speed is around 5–10 tok/s, so on these Macs a 3–4B model feels fine, an 8B model is usable for short answers, and a 14B model is slow. Measure your own Mac in Step 4; if you submit the result, it can become a community-measured row.

Step 1: Check your chip and macOS version

Click the Apple menu → About This Mac and read the Chip or Processor line:

Note two more lines on the same screen. macOS: Ollama needs 14 Sonoma or later. Memory: on an Intel Mac the model lives in ordinary system RAM next to macOS, so 16 GB is the practical minimum for 8B models; with 8 GB, stay at 1–4B and close other apps.

Step 2: Pick a CPU-friendly model

On CPU-only hardware, model size decides whether the experience is usable. Stick to small models at 4-bit:

As a rule of thumb on Intel: 3–4B is comfortable, 8B is the ceiling for daily use, and larger dense models are a wait. One exception if you have 32 GB or more (the 16-inch MacBook Pro and 2020 iMac can be configured with 64 and 128 GB): mixture-of-experts models such as GPT-oss 20B (3.6B active, about 13–14 GB) or Qwen 3.6-35B-A3B (3B active, about 20 GB) read about as much per token as a 3–4B dense model, so they run far faster than their size suggests. No Intel Mac run of either is published yet, so the index has no figure for them.

Step 3: Install Ollama or llama.cpp

Option A — Ollama (macOS 14 Sonoma or later)

Download Ollama from ollama.com, open the .dmg and drag it to Applications. Launch it once to install the command-line tool, then open Terminal and check:

ollama --version

Pull and run a small model:

ollama run llama3.2:3b

The first run downloads the model (about 2 GB), then opens a chat. On an Intel Mac Ollama runs on the CPU only. New to Ollama? The Install Ollama on Mac guide has the full walkthrough.

Option B — llama.cpp built from source (any recent macOS)

With Apple’s command-line tools and CMake installed, build a CPU-only copy (build instructions):

git clone https://github.com/ggml-org/llama.cpp
cd llama.cpp
cmake -B build -DGGML_METAL=OFF
cmake --build build --config Release -j 8

Download a 4-bit GGUF file of your model from Hugging Face, then chat with it, or serve it to other apps through an OpenAI-compatible API:

./build/bin/llama-cli -m path/to/model-Q4_K_M.gguf -ngl 0
./build/bin/llama-server -m path/to/model-Q4_K_M.gguf -ngl 0 --port 8080

-ngl 0 keeps every layer on the CPU. llama.cpp’s Metal backend is built for Apple Silicon, so do not count on the AMD or Intel graphics in an Intel Mac.

Step 4: Run it and measure your speed

Ollama prints its speed if you add --verbose:

ollama run llama3.1:8b --verbose

Ask a question and read the eval rate line at the end: that is decode speed in tokens per second. The prompt eval rate line is prompt processing, which on a CPU is also slow, so long pasted documents take a while before the first word appears. Compare your eval rate with the ceiling for your Mac in the table above.

An Intel Mac is useful for small, private, offline tasks: drafting text, summarizing a document, simple questions, light code help. It is not a machine for 30B-class dense models.

Step 5: Squeeze more speed out of an Intel Mac

Step 6: When to upgrade to Apple Silicon

If you use local models every day, Apple Silicon is a step change: its GPU shares one pool of unified memory with the CPU, with 1.6 to 7 times the memory bandwidth of these Intel Macs on the Macs compared here (68–307 GB/s), and up to about 29 times on an M5 Ultra. For an 8B model at 4-bit, the LLMCheck index estimates estimated~11 tok/s on a MacBook Air M1, estimated~25 tok/s on a 16 GB Mac mini M6 and estimated~49 tok/s on a Mac mini M5 Pro, against an upper estimate of about 7 tok/s on an i9 MacBook Pro 16-inch. An M1 is at least 1.6× faster; a current Mac mini is 3.5× to 7× faster or more.

The bigger change is what fits. A 24 GB Apple Silicon Mac runs Recursive Flint, the index’s pick for 24 GB, at estimated~71 tok/s on a Mac mini M6, and a 32 GB one runs Qwen 3.6-35B-A3B, tied for #1 in the index and its pick for 32 GB, at community59 tok/s. The Mac hardware guide and how much RAM you need cover the choice.

Entry pick: Mac mini M6 (16–32 GB)

The cheapest current way into Apple Silicon for local AI, from $899 with 16 GB at launch. Choose 24 or 32 GB if you can: that is the tier where the index’s best mixture-of-experts pick fits.

Check Mac mini M6 prices on Amazon →

Disclosure: As an Amazon Associate, LLMCheck earns from qualifying purchases. This doesn't affect the index’s figures or rankings.

Bottom line: keep your Intel Mac for small models and light, private tasks; it works. For capable local AI at reading speed and beyond, an Apple Silicon Mac is the upgrade that delivers it.