Mac LLM Fit

Given a model (Hugging Face URL, model card, or "{N}B parameters"), report the minimum Apple-Silicon Mac configuration that can run it locally at each quantization level.

Process

1. If the user gives a URL, fetch it with WebFetch and extract: parameter count, architecture, available quantizations, and file sizes.
2. If only a parameter count is known, estimate sizes from the table below.
3. Apply the memory math (see Rules) and pick the cheapest Mac that fits each quant.
4. Produce the output below. Keep it terse — this is a quick-check skill.

Output Format

**Model:** {name} — {N}B params, {architecture}

**Memory needs by quant:**

| Quant     | Weights size | Min unified memory | Notes               |
|-----------|--------------|--------------------|--------------------|
| Q4_K_M    | ~{x} GB      | {16/24/32/48/64} GB | {context budget}   |
| Q8_0      | ~{x} GB      | ...                 | ...                |
| BF16/FP16 | ~{x} GB      | ...                 | ...                |

**Cheapest viable Mac per quant:**
- Q4: {Mac model + RAM} — ~${price}
- Q8: {Mac model + RAM} — ~${price}
- BF16: {Mac model + RAM} — ~${price}

**Recommendation:** {one-line pick — which Mac for which use case}

**Tools to run it:** {LM Studio / Ollama / llama.cpp / MLX — pick what fits}

Rules

Memory math

• Weights size ≈ params × bytes-per-param. Bytes-per-param: BF16/FP16 = 2, FP8/Q8 = 1, Q5 ≈ 0.7, Q4 ≈ 0.55, Q3 ≈ 0.45.
• macOS reserves system memory; only ~67–75% of unified RAM is GPU-accessible by default. Add 6–8 GB headroom above weights for KV cache, context, and OS.
• Practical minimum unified memory ≈ weights_size × 1.5 (rounds up to next Mac SKU: 16 / 24 / 32 / 36 / 48 / 64 / 96 / 128 GB).
• For long context (>8k tokens) or batch inference, add another 4–8 GB.

Parameter-count quick table (Q4_K_M / Q8_0 / BF16 weights)

Mac SKU reference (current Apple Silicon lineup)

• Mac mini M4: 16 / 24 / 32 GB
• Mac mini M4 Pro: 24 / 48 / 64 GB
• MacBook Air M4: 16 / 24 / 32 GB
• MacBook Pro M4 Pro: 24 / 48 GB
• MacBook Pro M4 Max: 36 / 48 / 64 / 128 GB
• Mac Studio M4 Max / Ultra: up to 192 GB
• Pick the cheapest tier that meets the practical minimum. Prefer Mac mini for desk use, MacBook Pro if portability matters.

Style

• Be terse — the user wants a fast answer, not a teardown.
• If the model is too large for any reasonable Mac (e.g. 70B BF16 needs 192 GB), say so plainly and suggest a smaller quant.
• Mention Apple Silicon-native runtimes (MLX, llama.cpp Metal, LM Studio, Ollama) — don't recommend CUDA-only tools.
• Don't fabricate prices. Give a rough USD range or omit if uncertain.

Example

**Model:** Llama-3.1-8B-Instruct — 8B params, Llama 3 architecture

**Memory needs by quant:**

| Quant     | Weights size | Min unified memory | Notes                         |
|-----------|--------------|--------------------|-------------------------------|
| Q4_K_M    | ~5 GB        | 16 GB              | Fine for 4–8k context         |
| Q8_0      | ~8.5 GB      | 24 GB              | Comfortable headroom          |
| BF16      | ~16 GB       | 32 GB              | Needs M4 Pro tier+            |

**Cheapest viable Mac per quant:**
- Q4: Mac mini M4 16 GB — ~$599
- Q8: Mac mini M4 24 GB — ~$799
- BF16: Mac mini M4 Pro 48 GB — ~$1,999

**Recommendation:** Mac mini M4 24 GB is the sweet spot — runs Q8 with room for context.

**Tools to run it:** LM Studio or Ollama (both auto-pick Metal-accelerated builds).