Overview
An M5 Max MacBook Pro is the only MacBook Pro that Apple lists with 128GB of unified memory for a local LLM, and its 40-core GPU version has the highest bandwidth, 614GB/s; the M5 Pro tops out at 64GB and 307GB/s. The top configurations of both chips have an 18-core CPU (the 14-inch M5 Pro also has a 15-core CPU, 16-core GPU option). For a local model, memory size decides what loads and bandwidth is Apple’s spec for how fast memory can be read. What that means in tokens per second is unknown here, because this site publishes no speed figures and Apple publishes none for these Macs. Every spec below comes from Apple’s MacBook Pro specs, read on 2026-10-13.
How do the M5 Pro and M5 Max compare for a local LLM?
The M5 Max has more memory headroom and up to twice the bandwidth; the M5 Pro is the lower tier on both.
| Chip configuration | GPU cores | Memory bandwidth | Unified memory options |
|---|---|---|---|
| M5 Pro | 16 (15-core CPU, 14-inch) or 20 | 307GB/s | 24GB base on the 16-inch (14-inch base unknown); 48GB and 64GB options |
| M5 Max | 32 | 460GB/s | 36GB; 48GB, 64GB, 128GB options (see next section for which GPU) |
| M5 Max | 40 | 614GB/s | 48GB, 64GB, 128GB |
Source: Apple MacBook Pro specs. The plan for this page mentioned a 153GB/s tier. On Apple’s page that is the base M5 14-inch MacBook Pro (16GB, configurable to 24GB or 32GB), not Pro or Max, so it is outside this table; see apple-silicon-memory-bandwidth.
Apple’s launch text says “M5 Pro supports up to 64GB of unified memory” and “M5 Max supports up to 128GB of unified memory” (Apple newsroom). Apple says both chips feature “a new 18-core CPU with 6 super cores and 12 all-new performance cores” and “a Neural Accelerator in each GPU core” (same article). For what that hardware is, see m5-neural-accelerators.
Which memory option needs which chip?
Apple’s spec page ties each memory size to a chip in its configuration text:
- “36GB unified memory (M5 Max with 32-core GPU)”
- “48GB unified memory (M5 Pro and M5 Max with 40-core GPU)”
- “64GB unified memory (M5 Pro with 20-core GPU and M5 Max with 40-core GPU)”
- “128GB unified memory (M5 Max with 40-core GPU)”
So 128GB is an M5 Max with the 40-core GPU, which is also the 614GB/s version. Apple’s 128GB wording names only the 40-core GPU, so 128GB with the 460GB/s (32-core GPU) M5 Max is not confirmed by that text. Whether the lower M5 Pro GPU bin (16-core GPU, 15-core CPU) can be ordered with 64GB is unknown: Apple’s 64GB wording names only the “20-core GPU”. Check the configurator before you buy.
What fits in each memory size?
Fit is the site formula from methodology: weights_GB = params × bits / 8 / 1e9, plus kv_GB = 2 × layers × kv_heads × head_dim × context × 2 / 1e9, plus an assumed 2 GB overhead. The example is Qwen2.5-72B-Instruct: 72.7B parameters, and in config.json 80 layers, 8 KV heads, hidden size 8192 and 64 attention heads, so head_dim = 8192 / 64 = 128.
- Weights at 4.0 bits: 72.7 × 4 / 8 = 36.35 GB. At 8.0 bits: 72.7 GB.
- KV at context 4096: 2 × 80 × 8 × 128 × 4096 × 2 / 1e9 = 1.342 GB. At 32768 it is 8 times that, 10.737 GB.
| Case (Qwen2.5-72B) | Total GB | 36GB | 48GB | 64GB | 128GB |
|---|---|---|---|---|---|
| 4-bit, context 4096 | 39.692 | no | below | below | below |
| 4-bit, context 32768 | 49.087 | no | no | below | below |
| 8-bit, context 4096 | 76.042 | no | no | no | below |
| 8-bit, context 32768 | 85.437 | no | no | no | below |
“Below” means the total is under installed memory before the macOS reserve, which is unknown, so a close fit such as 49.087 GB in 64GB is unconfirmed in practice. The 4-bit and 8-bit columns use exactly 4.0 and 8.0 bits, not measured GGUF or MLX sizes. A 32B-class model at 4-bit and context 4096 totals 19.324 GB on how-much-ram-for-local-llm, under the 24GB base M5 Pro before the reserve. Picks by tier are on best-llm-for-64gb-mac and best-llm-for-128gb-mac. Coding models are on local-coding-llm-on-mac.
Is the M5 Max worth it over the M5 Pro for local LLMs?
If you need a model whose total is above 64GB, yes, because only the M5 Max offers 128GB. If your model total is under 64GB, memory size alone does not separate the two, and the difference is bandwidth (307 vs 460 or 614GB/s): 614 / 307 = 2.0. What that ratio does to speed is unknown. Apple’s only LLM figures are vendor claims: “up to 3.9x faster” LLM prompt processing for M5 Pro against a 16-inch M4 Pro (14-core CPU, 20-core GPU, 48GB), and “up to 4x faster” for M5 Max against a 16-inch M4 Max (16-core CPU, 40-core GPU, 128GB) (Apple newsroom). That is prompt processing, not generation, and Apple states its baseline; treat it as marketing until you measure with measure-tokens-per-second-mac. Memory ceilings per limit are on mac-gpu-memory-limit.
What is the MacBook Pro M5 LLM performance?
Unknown. This site publishes no tokens-per-second numbers and no primary source read for this page gives a generation speed for these Macs. Apple’s Neural Accelerator claim, “over 4x the peak GPU compute for AI compared to the previous generation”, is a vendor claim about peak compute, tested by Apple in February 2026 on preproduction 16-inch MacBook Pro systems, not a model speed.
Questions
Can an M5 Max MacBook Pro run a local LLM?
Yes, if the model total fits unified memory. The M5 Max is offered up to 128GB, and Qwen2.5-72B at 8-bit with context 32768 totals 85.437 GB by the formula above. The macOS reserve is unknown.
Is the M5 Pro enough for local LLMs?
For models whose total is under its memory, yes. The M5 Pro goes up to 64GB, and the 72B example at 4-bit and context 32768 totals 49.087 GB. Speed on that chip is unknown.
What is the difference between M5 Pro and M5 Max for local LLMs?
Memory ceiling and bandwidth. M5 Pro: up to 64GB, 307GB/s. M5 Max: up to 128GB, 460GB/s (32-core GPU) or 614GB/s (40-core GPU), all per Apple’s spec page.
What is the best local LLM for an M5 Max with 128GB?
This page does not rank models. It shows what fits: the 72B example fits at 8-bit with context 32768 (85.437 GB). Model picks are on best-llm-for-128gb-mac.
How fast is the MacBook Pro M5 for LLMs?
Unknown. Apple’s published LLM figures are vendor claims for prompt processing against older chips, not tokens per second. Measure your own with measure-tokens-per-second-mac.
Does 64GB of memory need the M5 Max?
No. Apple lists 64GB for an “M5 Pro with 20-core GPU” as well as the M5 Max with 40-core GPU. Whether the 16-core GPU M5 Pro can take 64GB is unknown.
Sources read on 2026-10-13
- Apple MacBook Pro specs: core counts, bandwidth per configuration, memory options and the chip-to-memory wording.
- Apple newsroom, MacBook Pro with M5 Pro and M5 Max: 64GB and 128GB ceilings, LLM prompt-processing claims (vendor claims). Prices there are not used.
- Apple newsroom, M5 Pro and M5 Max debut: memory and bandwidth ceilings, Neural Accelerator per GPU core, 4x peak GPU compute claim, February 2026 testing note.
- Qwen2.5-72B-Instruct card and its config.json: 72.7B parameters, 80 layers, 8 KV heads, hidden size 8192, 64 attention heads.
- methodology: weight, KV and 2 GB overhead formulas.