◆ MacModelFit

Local AI · Memory planner

Which Mac for which model?

Choose a model. See the memory it may need, and a capacity tier to investigate.

Your model

Includes the prompt and generated response. Longer context uses more memory.

Advanced model settings optional

Preset values use the linked publisher configuration. Llama 3.1 8B configuration source. Custom models need their own configuration.

For MoE, use total parameters, including inactive experts.

75% and 4 GB are adjustable planning assumptions. Apple’s recommended working set depends on the device; this page does not change your Mac’s settings.

Updates as you change inputs. All calculations stay in this browser.

Estimated model memory

—

Memory tier to investigate—

Estimate for one active text generation session.

Mac configurations to investigate

    How these examples are chosen

    These examples meet the recommended installed-memory tier. They do not establish model compatibility, response speed, current stock, or the best purchase for your budget. Start with the Mac you already have.

    Compare all memory tiers

    Memory tiers use GiB (1,073,741,824 bytes per GiB). These are generic memory tiers, not a list of configurations available for every Mac.

    Assumptions and limitations

    Weights = total parameters × effective bits per weight ÷ 8. KV cache = 2 × layers × KV heads × head dimension × allocated context × effective cache bits ÷ 8. Total includes the adjustable runtime allowance.

    GB uses 1,000,000,000 bytes; GiB uses 1,073,741,824. MLX allocation is rounded to 256 tokens. Quantized-cache estimates include group metadata. Actual files, activation peaks and runtime behavior may differ.

    Memory capacity does not establish speed, model quality or runtime compatibility. Hybrid, MLA, sliding-window and multimodal models need a separate estimate. This tool does not combine memory from separate Macs.

    Read the memory guide and sources →