On an Apple Silicon Mac with 8GB of unified memory, 84 of 224 catalog builds have good/excellent reference fits at 4K context, and 21 more run with trade-offs. The memory is shared with macOS and your other apps, so the headroom is smaller than the number suggests.
84
run well
21
with trade-offs
224
builds evaluated
A PC graphics card has its own memory, separate from system RAM. A Mac with Apple Silicon has one pool that the CPU and GPU share. That is why the answer differs from a 8GB graphics card, and why a model never gets the whole 8GB: macOS and open apps use some of it. These pages use a reference scenario. They do not know which M-series chip you have, and Apple sells several chips with the same memory size at very different speeds, so no speed is claimed here.
Editorial examples chosen for their documented tasks, all Q4_K_M at 4,096 tokens. Check the verdict before downloading. These are estimates, not benchmark rankings.
A small instruction-tuned model for trying local chat with more memory headroom. Review summaries against the source text.
Task and build source · Catalog checked 2026-09-13
Est. memory 2.9 GB · high confidence
Model fits within the unified memory GPU allocation.
ollama run llama3.2:3bAn 8B instruction-tuned alternative to compare on your own prompts. It needs more memory than the 3B example and can still give incorrect answers.
Task and build source · Catalog checked 2026-09-13
Est. memory 5.8 GB · high confidence
Model fits within the unified memory GPU allocation.
ollama run llama3.1:8bA code-focused instruct build. Review generated changes and run your tests; a memory fit is not evidence that its code is correct.
Task and build source · Catalog checked 2026-09-13
Est. memory 5.3 GB · high confidence
Model fits within the unified memory GPU allocation.
ollama run qwen2.5-coder:7bA 14B instruct build for comparing answers when memory allows. Larger size does not establish better results for every task.
Task and build source · Catalog checked 2026-09-13
Est. memory 10.2 GB · high confidence
Model does not fit within the unified memory GPU allocation.
ollama run qwen2.5:14bAfter opening Ollama's prompt, enter /set parameter num_ctx 4096 to match this calculation, within the model's supported maximum. Install a current Ollama release that supports the build. Setup and troubleshooting.
| Model | Build | Memory needed | Verdict |
|---|---|---|---|
| yi | yi:6b-v1.5-q4_K_MQ4_K_M | ~5.3 GB | Good fit |
| medgemma | medgemma:4bQ4_K_M | ~4.8 GB | Good fit |
| medgemma1.5 | medgemma1.5:4bQ4_K_M | ~4.8 GB | Good fit |
| translategemma | translategemma:4bQ4_K_M | ~4.8 GB | Excellent fit |
| qwen2.5vl | qwen2.5vl:3bQ4_K_M | ~4.7 GB | Excellent fit |
| phi4-mini-reasoning | phi4-mini-reasoning:3.8bQ4_K_M | ~4.6 GB | Excellent fit |
| ministral-3 | ministral-3:3bQ4_K_M | ~4.3 GB | Excellent fit |
| nemotron-3-nano | nemotron-3-nano:4bQ4_K_M | ~4.2 GB | Excellent fit |
| Phi-3 Mini | phi3:miniQ4_K_M | ~4.0 GB | Excellent fit |
| tev1 | tev1:4b-q4_K_MQ4_K_M | ~4.0 GB | Excellent fit |
| nemotron-mini | nemotron-mini:4bQ4_K_M | ~4.0 GB | Excellent fit |
| granite3-guardian | granite3-guardian:2bQ8_0 | ~4.0 GB | Excellent fit |
| qwen3-embedding | qwen3-embedding:4bQ4_K_M | ~3.7 GB | Excellent fit |
| phi4-mini | phi4-mini:3.8bQ4_K_M | ~3.7 GB | Excellent fit |
| phi3.5 | phi3.5:3.8b-mini-instruct-q4_K_MQ4_K_M | ~3.6 GB | Excellent fit |
| phi3 | phi3:3.8b-mini-4k-instruct-q4_K_MQ4_K_M | ~3.6 GB | Excellent fit |
| nuextract | nuextract:3.8b-q4_K_MQ4_K_M | ~3.6 GB | Excellent fit |
| llava-phi3 | llava-phi3:3.8bQ4_K_M | ~3.5 GB | Excellent fit |
| granite4.2 | granite4.2:3bQ4_K_M | ~3.4 GB | Excellent fit |
| cogito | cogito:3bQ4_K_M | ~3.4 GB | Excellent fit |
| granite-code | granite-code:3b-base-q4_K_MQ4_K_M | ~3.2 GB | Excellent fit |
| smallthinker | smallthinker:3b-preview-q4_K_MQ4_K_M | ~3.2 GB | Excellent fit |
| granite4 | granite4:3bQ4_K_M | ~3.2 GB | Excellent fit |
| granite4.1 | granite4.1:3bQ4_K_M | ~3.2 GB | Excellent fit |
| hermes3 | hermes3:3bQ4_K_M | ~3.1 GB | Excellent fit |
Showing the 25 largest of 84 models that run well.
Each model is evaluated at 4K context against 8GB of shared memory on a reference Apple Silicon Mac running the Metal backend, counting weights, KV cache, compute buffers and runtime overhead. The engine reserves memory for macOS and estimates the share available to the GPU. That share is a heuristic. Available memory and current memory pressure are unknown; check Activity Monitor during a real run. Ollama requires macOS 14 or later. See the 16 GB Mac walkthrough for a worked example, or the scan to measure your own Mac.