Capability frontier

The right Mac for local AI starts at a specific tier.

Choose the workload and what you optimize for. Silicon Score shows the first Apple Silicon tier that is genuinely viable, what ceiling it unlocks, and whether that read is measured or still directional.

28 Mac configs 18 frontier-priority models Frontier current through February 27, 2026

Open buy answer Open run answer Plan fleet Check worth

Workload

Optimize for responsive daily-driver use with enough room for stronger local assistants.

Optimize for

Push toward the smartest model you can still use productively.

Serious tier now

24GB

Mac Mini M4 24GB · $599

Current frontier

Qwen3.5-27B

Qwen 3 32B, Qwen3.5-27B, Devstral Small 2 24B

Live answer

24GB is where coding starts to feel properly viable.

Mac Mini M4 Pro 24GB is the strongest current directional read here with Qwen3.5-27B, but tier-specific throughput still needs measurement. 32GB changes the story like this: Mostly more headroom and better confidence than the tier below.

Directional

What is visible now

Directional field signal

Frontier movement is visible here from field signals, but tier-specific throughput still needs direct measurement.

Current ceiling

Qwen3.5-27B

Mac Mini M4 Pro 24GB currently anchors this tier for coding, but direct speed measurement is still thin.

Throughput read

Needs direct measurement

Current tok/s reads here are directional field reports rather than tier-specific benchmark evidence.

Local cost read

Hold until measured

Do not compare local cost across tiers until speed is measured on this hardware class.

Why this tier

Ceiling jump

Moves the ceiling from Qwen3.5-9B to Qwen 3 32B.

Below this tier

16GB

16GB tops out around Qwen3.5-9B.

Frontier movement

Qwen 3 32B, Qwen3.5-27B, Devstral Small 2 24B

Recent frontier movement here centers on Qwen3.5-27B and Devstral Small 2 24B.

Evidence state

Directional

Useful for frontier movement, but not settled by direct benchmark coverage yet.

Primary confidence

Low confidence

Speed is directional from Apple Silicon field reports. Runtime mentions cluster around MLX.

Throughput read

Needs direct measurement

Current tok/s reads here are directional field reports rather than tier-specific benchmark evidence.

Local cost read

Hold until measured

Do not compare local cost across tiers until speed is measured on this hardware class.

Open this tier Check entry-tier cost Audit evidence

Capability ladder

Where the answer actually changes.

Each row is a memory tier. The point is not more RAM in the abstract, but a different practical frontier, a different confidence level, or a different speed class.

Directional

Directional field signalUseful for frontier movement, but not settled by direct benchmark coverage yet.

Tier read

Mac Mini M4 Pro 24GB is the strongest current directional read here with Qwen3.5-27B, but tier-specific throughput still needs measurement.

Recent frontier movement here centers on Qwen3.5-27B and Devstral Small 2 24B.

Machine spread

Mac Mini M4 24GB to Mac Mini M4 Pro 24GB

No measured speed leader is settled yet. The quickest current directional read is Mac Mini M4 Pro 24GB at 22.0 tok/s.

Confidence

Low confidence

Speed is directional from Apple Silicon field reports. Runtime mentions cluster around MLX.

Power band

~25-45W

Mac Mini M4 Pro 24GB currently anchors this tier for coding, but direct speed measurement is still thin.

Frontier optionSpeedCostQuantEvidence

Fastest daily driver

Qwen 3 32B

Mac Mini M4 24GB

22.0 tok/s

Hold until measured

Q4_K_M

Benchmark estimate

Smartest viable model

Qwen3.5-27B

Mac Mini M4 Pro 24GB

8.5 tok/s

Hold until measured

Q5_K_M

Field report

Another viable option

Devstral Small 2 24B

Mac Mini M4 24GB

—

Hold until measured

Q6_K

Fit estimate

Entry tier

Mac Mini M4 24GB

$599

Best machine

Mac Mini M4 Pro 24GB

Mac Mini M4 Pro 24GB is the strongest current directional read here with Qwen3.5-27B, but tier-specific throughput still needs measurement.

Ceiling move

Qwen3.5-27B

Moves the ceiling from Qwen3.5-9B to Qwen 3 32B.

Inspect best machine Check local cost

Confidence posture

Measured and directional answers stay separate.

The surface should show where the answer is benchmark-backed, where it is estimated from nearby coverage, and where it is still primarily a frontier signal.

Measured

Benchmark-backed answer

Direct benchmark coverage is available on this hardware class, so the answer is materially firmer.

Estimated

Estimated from nearby benchmarks

Anchored by nearby benchmark coverage, but not a direct machine-and-quant match.

Directional

Directional field signal

Useful for frontier movement, but not settled by direct benchmark coverage yet.

Fit-first

Fit is clearer than speed

Sizing math is usable now, but throughput still needs direct measurement.

Anchor lines

Keep one firm anchor and one open watch in view.

A useful decision surface should show the most settled desktop path and the highest-interest open question at the same time.

Measured desktop anchor

Mac Studio M3 Ultra 256GB

Fit-first

Qwen3.5-27B · $7,499

Qwen3.5-27B is the smartest viable model for coding on this Mac. Speed coverage is not integrated yet.

Open this Mac Check worth

Buyer watch

M5 Max 128GB track

Current public anchors are 8B 61.6 tok/s and 14B 34.3 tok/s on a 36GB M5 Max. Keep the 128GB buyer interest visible, but do not publish a settled 128GB machine answer until direct coverage lands.

Open M5 guide Audit current refs