Key fact: ollama run qwen3.8:27b needs about 18.2 GB at Q4 before macOS takes its share of unified memory.
How we got it: 27 × 0.5 × 1.2 + 2 = 18.2. Ollama listed an 18 GB download on 2026-10-03, plus a qwen3.8:27b-mlx tag.
Checked:
By LLM Hardware team · Methodology ·
Editorial: AI drafted the step-by-step; the Apple-specific gotchas (Metal toggles, MLX vs Ollama tradeoffs) come from the MLX, Ollama and llama.cpp sources listed at the end of this guide.
Checked 2026-10-13. Author: LLM Hardware team. Amazon prices are not stored or shown on this site.
Updated 10 November 2026 · Ollama, MLX, and the Qwen3.8 / Gemma 4 tags · Apple Silicon and Intel
Key fact: ollama run qwen3.8:27b needs about 18.2 GB at Q4 before macOS takes its share of unified memory.
How we got it: 27 × 0.5 × 1.2 + 2 = 18.2. Ollama listed an 18 GB download on 2026-10-03, plus a qwen3.8:27b-mlx tag.
Checked:
Key fact: Gemma 4 E4B is a 16 GB Mac model. Gemma 4 26B-A4B is a 36 GB Mac model.
How we got it: Unsloth E4B Q4_K_M file 4.98 GB. 26B-A4B formula on 25.2B total is 17.1 GB, and a 24 GB Mac is shared with the OS.
Checked:
Methodology explains the VRAM formula, the bytes per weight at each quantization, and the full source list.
Apple Silicon Macs from the first M1 models to the current M5 and M6 models run open-weight LLMs locally with Ollama, LM Studio or MLX, and Ollama also runs on Intel Macs on the CPU only. All three tools need macOS 14 Sonoma or newer. Older M1 to M4 Macs use the same steps; what changes is how much unified memory you have, which sets the models that fit. On 10 November 2026 Apple listed 16 GB to 32 GB for M6, 24 GB to 64 GB for M5 Pro, 36 GB to 128 GB for M5 Max and 96 GB to 512 GB for M5 Ultra. Choosing a new Mac? See Apple Silicon for LLMs. On a PC, see How to Run LLMs on Windows 10 and 11: Ollama and LM Studio.
Requires macOS 14 Sonoma or newer. Step 1 uses Homebrew; if you do not use Homebrew, download the app from ollama.com instead, which is the install method the Ollama macOS docs prefer. Download time depends on the model size and your connection.
Step 1: Install Ollama
brew install ollama
Alternative: curl -fsSL https://ollama.com/install.sh | sh or download from ollama.com
Step 2: Start the Ollama server
ollama serve
Leave this running in one Terminal tab. The server listens on http://localhost:11434
Step 3: Download and run a model
ollama run qwen3:8b
Qwen3 8B at Q4_K_M is 8 × 0.5 × 1.2 + 2 = 6.8 GB. That is a 16 GB Mac once the OS is counted, and a squeeze on 8 GB. phi4:14b is 10.4 GB. llama3.3:70b is 44 GB.
On Apple Silicon: Metal GPU is automatic
Ollama detects an M-series chip and uses Metal with no extra flags. Unified memory is one pool shared with macOS, so a 24 GB Mac is not the same fit as a 24 GB discrete GPU.
These tags were on the Ollama library on 3 October 2026 and were still listed on 10 November 2026. The gigabytes in the right column are this site’s short-context formula, not a speed figure. The fit table below uses the same formula.
32 GB or larger Mac, or a 24 GB Mac if you keep the chat short
ollama run qwen3.8:27b
27B dense, Apache-2.0. Formula 18.2 GB. The library listed an 18 GB download and a qwen3.8:27b-mlx tag. Unsloth’s UD-Q4_K_XL file was 17.6 GB.
Same machines, MoE instead of dense
ollama run gemma4:26b
Google’s card: 25.2B total, 3.8B active, Apache-2.0. Formula on the total is 17.1 GB. Ollama listed 16 to 19 GB. There is a gemma4:26b-mlx tag.
16 GB Mac, edge multimodal
ollama run gemma4:e4b
Unsloth’s E4B Q4_K_M file is 4.98 GB. E2B’s Q4_K_M file is 3.11 GB (ollama run gemma4:e2b). Both include text, image, and audio on Google’s card.
Do not run Kolibri-1 here. Aleph Alpha’s card lists about 78 GB of FP8 weights and datacenter GPUs. Buy the memory before the weights: the in-stock Amazon Macs are on the Apple Silicon guide, and the arithmetic is on the VRAM calculator. If a discrete GPU is the better fit, use the GPU guide.
Each cell is a fit verdict from this site’s VRAM formula at Q4_K_M, not a speed figure. Apple Silicon uses Metal; Intel uses CPU only.
| Mac | Unified Memory | Qwen3 8B (6.8 GB) | Phi-4 14B (10.4 GB) | Llama 3.3 70B (44 GB) | Notes |
|---|---|---|---|---|---|
| Any Intel Mac | System RAM | CPU, needs 6.8 GB | CPU, needs 10.4 GB | CPU, needs 44 GB | CPU-only inference: slow but works for small models |
| M1/M2/M3 8 GB | 8 GB unified | ✗ (4 GB usable) | ✗ | ✗ | About 4 GB usable after the macOS reserve: small models only |
| M1/M2/M3/M4 16 GB | 16 GB unified | fits | fits | ✗ | Fits 8B and 14B at Q4 with a short context |
| M4 Mac Mini 24 GB | 24 GB unified | fits | fits | ✗ | Prior generation. Same fit as any 24 GB Mac |
| M4 Pro 24 GB | 24 GB unified | fits | fits | ✗ | Same memory fit as the M4 24 GB; the Pro chip adds memory bandwidth |
| M4 Pro 48 GB | 48 GB unified | fits | fits | tight | Prior generation. 70B at Q4 is 44 GB against 44 GB usable: tight, short context only |
| M4 Max 64 GB | 64 GB unified | fits | fits | fits | Prior generation. Safer size for 70B at Q4; longer context needs more |
| M5 Pro 24 GB | 24 GB unified | fits | fits | ✗ | Current Mac mini option. 27B at Q4 (18.2 GB) is tight |
| M5 Max 36 GB | 36 GB unified | fits | fits | ✗ | Room for 27B to 32B at Q4. Not 70B |
| M6 16 GB | 16 GB unified | fits | fits | ✗ | Base memory on the current M6 Mac mini. 8B and 14B at Q4, short context |
| M6 24 or 32 GB | 24 or 32 GB unified | fits | fits | ✗ | More room for context. 32B at Q4 (21.2 GB) fits 32 GB, not 24 GB |
| M5 Pro or M5 Max 48 GB | 48 GB unified | fits | fits | tight | 70B at Q4 is 44 GB against 44 GB usable: short context only |
| M5 Pro or M5 Max 64 GB | 64 GB unified | fits | fits | fits | Safer size for 70B at Q4; longer context needs more |
| M5 Max 128 GB or M5 Ultra | 96 to 512 GB unified | fits | fits | fits | Room for 70B with longer context, or larger models |
Fits = the Q4_K_M estimate from the VRAM calculator (params × 0.5 × 1.2 + 2 GB) leaves at least 10% of the Mac’s memory free after a 4 GB macOS reserve. Tight = it fits inside the last 10%. ✗ = it does not fit at Q4_K_M. Intel Macs run on the CPU from system RAM, so their cells show the memory the model needs. Buy-now prices and stock are in the Apple Silicon guide.
Memory options on Apple tech specs pages, checked 10 November 2026: Mac mini with M6 (16, 24 or 32 GB) or M5 Pro (24, 48 or 64 GB); Mac Studio with M5 Max (36, 48, 64 or 128 GB) or M5 Ultra (96, 256 or 512 GB). Not every size is offered with every chip configuration. The 4 GB macOS reserve in the rule above is this site's planning convention, not an Apple figure. The chip options and a fit table for each memory size are in the Apple Silicon guide.
Four tools cover most Mac LLM use cases. Ollama is the foundation most others build on. Even if you plan to use a GUI, install Ollama first.
| Tool | Type | Install | Best For |
|---|---|---|---|
| Ollama | CLI + API server | brew install ollama | Developers, scripting, API access |
| LM Studio | GUI app | lmstudio.ai download | Beginners, model browsing |
| Open WebUI | Browser UI (requires Ollama) | pip install open-webui | ChatGPT-like interface |
| MLX Community | Python framework | pip install mlx-lm | Developers, Apple Silicon native speed |
Adding a ChatGPT-style interface
Open WebUI runs in your browser and connects to the Ollama server. Install it with pip, then open localhost:8080.
pip install open-webui open-webui serve
MLX on Apple Silicon
MLX is an array framework for machine learning on Apple silicon from Apple machine learning research. mlx-lm runs LLMs on it and needs macOS 14 or newer with a native Python 3.10 or newer. LM Studio can also load MLX models on Apple Silicon.
pip install mlx-lm mlx_lm.generate --model mlx-community/Qwen3-8B-4bit \ --prompt "Explain unified memory"
Metal GPU acceleration is automatic
Ollama detects Apple Silicon at startup and uses Metal for all GPU operations. No flags, environment variables, or configuration needed. You can confirm GPU usage in Activity Monitor: open it, go to Window > GPU History, and watch GPU utilization spike to 80-100% while a model generates.
Unified memory is one shared pool
On a PC with a discrete GPU, the model has to fit the card's own VRAM. On Apple Silicon, the GPU uses the same unified memory as macOS and your apps, so a 64 GB Mac does not give the model all 64 GB. The fit table on this page holds back 4 GB for macOS, which is this site's planning convention, not an Apple guarantee; browsers and other open apps take more.
Memory pressure and swap
If you load a model that is too large for available memory, macOS swaps to the SSD and generation slows sharply. Watch Memory Pressure in Activity Monitor: if it climbs while a model loads, use a smaller model or a smaller quantization. By the site formula, Qwen3 8B at Q4_K_M is 6.8 GB, a squeeze on an 8 GB Mac and comfortable on 16 GB. A 14B model at Q4_K_M is 10.4 GB and fits a 16 GB Mac at a short context.
Check memory pressure: Activity Monitor > Memory tab > Memory Pressure (bottom of window).
Choosing quantization on Apple Silicon
Q4_K_M is the standard starting point. It balances size and quality well. On 16 GB Macs, stick to Q4_K_M for 7B models. On 24 GB, you can run 7B at Q8 (near-lossless) or 14B at Q4_K_M. The M4 Pro chip at 48 GB can run 13B at Q8 and 34B at Q4_K_M comfortably. Higher quantization (Q8, FP16) improves output quality at the cost of more memory.
Ollama runs on Intel Macs using CPU-only inference, on macOS 14 Sonoma or newer. Intel integrated graphics do not have Metal support that Ollama uses for LLM acceleration, so all computation falls on the CPU cores. Expect generation to be noticeably slower than on Apple Silicon.
What works on Intel Mac
What to avoid on Intel Mac
Recommendation for Intel Mac users
Install Ollama and run ollama run qwen3:8b to test. If the speed is acceptable for your use case, great. If you want faster inference, an Apple Silicon Mac mini runs the same models on its GPU through Metal.
Error: this model requires mlx support, but the mlx runtime is not available
Ollama's published docs do not define this exact error string (checked 12 November 2026 against docs.ollama.com). Official release notes and install docs do say when the MLX runtime is used and which packages include it. Since Ollama v0.40.0, "on Apple Silicon devices, model architectures supported by the MLX runtime will automatically run on MLX." The macOS docs prefer installing the official .dmg app (macOS Sonoma or newer; Apple M series get CPU and GPU support; Intel Macs are listed as CPU only). If a model needs the MLX runtime and this install does not have that runtime available, the pull or run fails with the error above.
Fix on a Mac: install or update the official Ollama app from ollama.com using the .dmg, on Apple Silicon, so you are on a current official release that includes the MLX runtime (v0.40.0 or later for the default described above). Then retry the model. Some models were limited to Apple hardware in a given release. The v0.40.12 notes said lfm3 and d1 2b were "supported on Apple hardware in this release. Support for Linux/Windows coming soon." The v0.40.13 notes then said LFM3 models "now run on Linux and Windows as well as Apple silicon." If the error appeared for one of those models on an older official release, use a current official release.
If you hit the same error on Windows with the standalone CLI, the Windows docs list an additional package, ollama-windows-amd64-mlx.zip, under "MLX (CUDA)," to extract into the same directory as the standalone CLI. Official Linux install docs describe the standard and ROCm tarballs but do not currently document the MLX Linux package, so this page does not give Linux extract steps.
Error: model requires more system memory than available
The model's Q4_K_M weights exceed your available RAM minus what macOS needs. Fix: run a smaller model or use a more aggressive quantization.
# Instead of phi4:14b on an 8 GB Mac, use: ollama run qwen3:8b # Or pull the 4-bit version explicitly: ollama pull qwen3:8b-q4_K_M
Model loads but generation is extremely slow
macOS is swapping model weights to the SSD. The model is too large for available RAM. Open Activity Monitor, check Memory Pressure. It is likely orange or red. Fix: close other applications, restart Ollama, or switch to a smaller model.
# See each loaded model, its size and the PROCESSOR split: ollama ps # Unload a model to free memory: ollama stop qwen3:8b
ollama: command not found after brew install
Homebrew's bin directory is not in your PATH. Run the path fix or use the direct installer instead.
# Add Homebrew to PATH (add to your ~/.zshrc too): export PATH="/opt/homebrew/bin:$PATH" # Or use the direct installer: curl -fsSL https://ollama.com/install.sh | sh
Model stuck at "pulling manifest" or download stalls
Network issue or Ollama registry timeout. Press Ctrl+C to cancel, then retry. If it keeps stalling, try pulling a different model first to confirm the registry is reachable, or check ollama.com status. Large models (70B) can take 30+ minutes on a typical home connection. The progress bar may appear frozen but is still downloading.
Three options: (1) Download the macOS app from ollama.com, open the .dmg and drag Ollama to Applications, which is the method the Ollama macOS docs prefer. (2) Run "curl -fsSL https://ollama.com/install.sh | sh"; the official script supports macOS and installs the same app. (3) Run "brew install ollama" if you already use Homebrew. Then run "ollama run qwen3:8b" to download and run your first model. If the app is not running, "ollama serve" starts the server. Ollama needs macOS 14 Sonoma or newer.
Yes. Ollama supports GPU acceleration on Apple devices through the Metal API, and its macOS docs list CPU and GPU support on Apple M series chips. No configuration is needed. To check, run ollama ps while a model is loaded: the PROCESSOR column shows 100% GPU when the whole model is on the GPU. On Intel Macs, Ollama runs on the CPU only, which is much slower.
8 GB unified is a small-model machine once macOS has taken its share. Qwen3 8B at Q4_K_M is 8 × 0.5 × 1.2 + 2 = 6.8 GB by the site formula, so 16 GB is the comfortable Mac for that tag. A 27B model at Q4 (18.2 GB) is tight on 24 GB and fits 32 GB or more. Llama 3.3 70B at Q4_K_M is a 44 GB estimate at short context, so 48 GB is tight and 64 GB is the safer size; longer context and other open apps need more. These fits use the 4 GB macOS reserve this site assumes, a planning convention rather than an Apple guarantee, and the sizes are formula estimates, not measured allocations.
Yes, with limits. The Ollama macOS docs list Intel (x86) Macs as CPU only, and Ollama needs macOS 14 Sonoma or newer, so an Intel Mac that cannot run Sonoma is out. LM Studio does not support Intel-based Macs, and MLX requires Apple silicon. On the CPU, stick to small models, about 8B and below at Q4, for experiments. An Apple Silicon Mac runs the same models on its GPU through Metal.
On 10 November 2026 the Ollama library listed ollama run qwen3.8:27b at 18GB, with a qwen3.8:27b-mlx tag, and ollama run gemma4:26b at 16GB to 19GB, with a gemma4:26b-mlx tag. Our VRAM formula is 18.2 GB for the 27B Q4 and 17.1 GB for Gemma 4 26B-A4B counted at 25.2B total parameters. Use those on a 32 GB or larger Mac, or on 24 GB only with a short context. On a 16 GB Mac use ollama run gemma4:e4b or gemma4:e2b instead. Do not point qwen3:32b at Qwen3.8, and do not expect either 27B-class model to fit an 8 GB Mac.
MLX is an array framework for machine learning on Apple silicon from Apple machine learning research, and mlx-lm runs LLMs on it. Install it with "pip install mlx-lm" (MLX needs macOS 14 or newer and a native Python 3.10 or newer) and load MLX-format models from Hugging Face. Ollama with Metal is the quickest to get running, and the Ollama library also lists -mlx tags for some models, such as qwen3.8:27b-mlx. This page does not publish speed comparisons between MLX and Ollama, so compare them on your own Mac if speed matters.
Behaviour, file-format and runtime details on this page are pulled from primary upstream docs and the sources listed below. The full sitewide methodology lives on the methodology page. For this guide we relied most on:
Spot a number that does not match the linked source? Contact us and we will update the guide.