Run on your Mac

Run Outlier Plus on Mac Studio M4 Ultra

Last updated 2026-06-18 · Outlier v1.11.469

What does Outlier Plus run like on Mac Studio M4 Ultra?

The pairing under examination is Outlier Plus (209 GB MLX 4-bit, Qwen3.5-397B-A17B base, 32K default context, 64 GB unified memory minimum) on a Mac Studio M4 Ultra (Apple M4 Ultra, 28-32 CPU cores, 60-80 GPU cores, 64|128|192|256|512 GB unified memory, 1092 GB/s, 2025). Outlier loads this tier as one mlx_lm process inside the bundled FastAPI sidecar. Apple Silicon decode on dense 4-bit weights is bandwidth-bound, so the ceiling on this exact Mac Studio M4 Ultra scales as the ratio 1092 / 800 = 1.36× the published M1 Ultra number.

How much memory does the model take during generation?

Outlier’s Outlier Plus has a measured peak generation footprint of about 14.04 GB. The Mac Studio M4 Ultra base configuration ships with 64|128|192|256|512 GB unified, and the rule of thumb is to leave roughly 4 GB for the OS and one open browser tab. On Mac Studio M4 Ultra the Outlier Plus bandwidth-scaled estimate is ~2.2 tok/s [estimated from family-bandwidth ratio]; derivation: 1.59 tok/s on M1 Ultra × (1092 / 800) bus ratio for the Apple M4 Ultra. Treat as a first-order projection — the Outlier Plus number on this 2025 machine has not been formally measured.

What is the install path on Mac Studio M4 Ultra for Outlier Plus?

The 2025 Mac Studio M4 Ultra ships with 64|128|192|256|512 GB of unified memory, so headroom for the 64 GB Outlier Plus requirement is 0 GB on the base SKU. Step-by-step install instructions live on the install guide; the part that varies for this Apple M4 Ultra machine is that the 209 GB pull from mlx-community/Qwen3.5-397B-A17B-4bit over HTTPS lands in ~/Library/Application Support/Outlier/models/plus/ and download time is bandwidth-limited by the network, not by the 28-32-core Apple M4 Ultra CPU.

What context window can I use on a 64 GB Mac?

The Outlier Plus tier defaults to 32K context and caps at 256K. KV cache scales linearly with context length on dense models, so longer contexts trade headroom for capacity. On a 64|128|192|256|512 GB Mac Studio M4 Ultra, the default context is the safe starting point.

What works well on this pairing, and what is still rough?

• The Mac Studio M4 Ultra (Apple M4 Ultra, 1092 GB/s) is bandwidth-bound on dense 4-bit decode; the throughput ceiling for Outlier Plus on this exact machine is approximately 1092 divided by 209 GB tokens per second.
• The Mac Studio M4 Ultra chassis dissipates heat well; sustained decode at the Outlier Plus tier did not show thermal throttling in our 2026-05 internal soak.
• The Apple M4 Ultra (1092 GB/s, 60-80 GPU cores) is current Apple Store silicon as of 2025; new-purchase tok-per-second-per-dollar is best in the lineup at the Outlier Plus tier.
• Cold boot of Outlier on a 28-32-core Apple M4 Ultra is roughly 50 seconds while the bundled Python.framework warms; this is documented on the v1.9 backlog under the boot-time investigation thread.

What is the unique number for Outlier Plus on Mac Studio M4 Ultra?

209 GB on disk against 64|128|192|256|512 GB unified memory means the weights alone consume about 327% of the base SKU’s RAM on the Mac Studio M4 Ultra.

Should I pick a different tier?

For lighter-weight work on this Apple M4 Ultra, the Code tier is the next step down and runs visibly faster. The unified-memory explainer works the bandwidth math out, and the bandwidth ratio for Mac Studio M4 Ultra is 1.36× M1 Ultra reference.

Download Outlier for Mac

Requires Apple Silicon (M1, M2, M3, or M4) — Intel Macs are not supported. macOS 12+.