mac-ai-doctor 0.1.2

Transparent unified-memory fit estimates for local AI models

🩺 Mac AI Doctor

Will this AI model fit in your Apple Silicon Mac's unified memory?

maid reads small metadata files—never model weights—and estimates weight, KV-cache, and runtime memory as a range, then gives a plain verdict.

$ maid check ibm-granite/granite-switch-4.1-3b-preview --memory-gb 16

     ibm-granite/granite-switch-4.1-3b-preview
┌────────────┬────────────────┐
│ Weights    │        8.80 GB │
│ KV cache   │        0.34 GB │
│ Runtime    │        1.31 GB │
│ Peak range │ 11.49–13.06 GB │
│ Available  │        16.0 GB │
└────────────┴────────────────┘
╭───────────────────── TIGHT · high confidence ──────────────────────╮
│ May fit, but close memory-heavy apps or reduce context/concurrency. │
╰─────────────────────────────────────────────────────────────────────╯

Install

Requires uv.

uv tool install mac-ai-doctor

Use

maid system                                              # your chip and memory
maid check meta-llama/Llama-3.1-8B-Instruct              # a Hugging Face model
maid check ~/Models/model-q4.gguf --context 8192         # a local GGUF
maid compare org/model-a org/model-b --concurrency 2     # side by side
maid recommend --memory-gb 16 --task coding              # a starting point
maid check MODEL --json                                  # machine-readable

Accepts a Hugging Face ID, a local .gguf file, or a local/MLX directory with config.json.

Options: --memory-gb N (required off macOS), --context, --concurrency, --kv-dtype (fp32/fp16/bf16/int8/q8), --json. Tasks: coding, chat, vision.

Verdicts

Verdict	Meaning
COMFORTABLE	High estimate ≤ 80% of memory.
TIGHT	Fits, but leaves < 20% headroom.
UNLIKELY	Exceeds memory.
UNKNOWN	Weight size unavailable.

Confidence reflects metadata completeness, not prediction accuracy.

What it reads

Only bounded metadata: config.json, the safetensors index, and the model API's file-size listing—or the 24-byte header of a local GGUF. Weight contents are never downloaded. No credentials are collected.

How it estimates

Decimal GB. This is screening, not a benchmark—verify with your actual runtime. No tokens/second prediction is made.

weights = stored_weight_bytes × 1.06
KV      = 2 × layers × KV_heads × head_dim × context × concurrency × dtype_bytes
runtime = max(1 GB, weights × 12%) + 0.25 GB × concurrency
range   = subtotal × 1.10 .. subtotal × 1.25

Unified memory is shared with macOS, apps, and the GPU. Memory mapping can help; multimodal image encoders may add memory not shown in text config metadata.

Troubleshooting

• Can't detect memory — pass --memory-gb N (auto-detection needs macOS).
• 401/403 or gated — authenticate, or point at a downloaded config.json.
• No weight size — use a repo with a safetensors index/API sizes, or a local GGUF (v2/v3).

Develop

git clone https://github.com/barvhaim/mac-ai-doctor.git && cd mac-ai-doctor
uv sync --group dev
uv run ruff check . && uv run mypy src && uv run pytest

See CONTRIBUTING.md. MIT licensed.