Overview
The M5 Ultra is the Mac Studio chip with the most unified memory Apple lists for local LLMs: 96GB as the base, configurable to 256GB or 512GB, with 1.2TB/s of memory bandwidth, according to Apple’s Mac Studio specs. Apple’s September 22 announcement says the 512GB configuration is “coming in late October”. Speed in tokens per second is unknown here; no primary source gives a figure and this site does not publish its own. Measure it yourself with measure-tokens-per-second-mac.
What are the M5 Ultra and M5 Max Mac Studio specs?
These rows are copied from Apple’s Mac Studio specs, fetched 2026-10-13.
| Chip | CPU | GPU | Neural Engine | Unified memory | Memory bandwidth |
|---|---|---|---|---|---|
| M5 Max (base) | 18-core | 32-core | 16-core | 36GB | 460GB/s |
| M5 Max (configurable) | 18-core | 40-core | 16-core | 48GB, 64GB or 128GB | 614GB/s |
| M5 Ultra (base) | 30-core | 64-core | 32-core | not tied to a row on the page | 1.2TB/s |
| M5 Ultra (configurable) | 36-core | 80-core | 32-core | 96GB base, 256GB or 512GB | 1.2TB/s |
The spec page lists 512GB as the maximum memory and leaves the price blank. The August announcement gives starting prices of 5,499 (M5 Ultra), US. Prices for memory upgrades are unknown.
What is the M5 Ultra memory bandwidth, and how does it compare with M3 Ultra?
Apple lists 1.2TB/s for M5 Ultra and calls it “50 percent more than M3 Ultra” in its M5 Ultra announcement. The 2025 Mac Studio spec page lists M3 Ultra at 819GB/s (Apple Support 122211).
| M3 Ultra (2025) | M5 Ultra (2026) | |
|---|---|---|
| Memory bandwidth | 819GB/s | 1.2TB/s |
| Base memory | 96GB | 96GB |
| Largest memory, per Apple Support 122211 | 256GB | not listed there |
| Largest memory, per the 2025 newsroom | 512GB | 512GB (spec page) |
Apple’s own pages disagree on M3 Ultra: the support page lists 96GB configurable to 256GB, while the March 2025 newsroom says “starts with 96GB of unified memory, which can be configured up to 512GB”. This site does not resolve that conflict. Check the configurator before buying a used or discounted M3 Ultra.
Vendor claims, tested by Apple and not by this site: M5 Ultra has “up to 4.5x the peak GPU compute for AI compared to M3 Ultra”, measured with “select industry-standard benchmarks” that Apple does not name, and “up to 4x faster” LLM prompt processing than M3 Ultra in LM Studio (the 4.5x figure is in the M5 Ultra announcement, the 4x figure in the Mac Studio announcement, both linked in Sources). Prompt processing is not generation speed, and Apple gives no generation figure on those pages.
What fits on an M5 Max 128GB versus an M5 Ultra 96GB?
Both have 96GB or more. Use the formulas on methodology: weights_GB = params x bits / 8 / 1e9, kv_GB = 2 x layers x kv_heads x head_dim x context x 2 bytes / 1e9, and total_GB = weights + kv + 2 GB overhead (the overhead is an assumption, not a measurement).
Worked example, Qwen2.5-72B-Instruct: the card says 72.7B parameters; its config.json has 80 layers, 8 KV heads, hidden size 8192 and 64 attention heads, so head_dim = 8192 / 64 = 128.
| Setting | Weights | KV cache | Overhead | Total |
|---|---|---|---|---|
| 4-bit, context 4096 | 36.35 GB | 1.342 GB | 2 GB | 39.692 GB |
| 4-bit, context 32768 | 36.35 GB | 10.737 GB | 2 GB | 49.087 GB |
| 8-bit, context 32768 | 72.7 GB | 10.737 GB | 2 GB | 85.437 GB |
KV at 32768 is 2 x 80 x 8 x 128 x 32768 x 2 / 1e9 = 10.737 GB. Every row is below 96GB of installed memory, so a 96GB M5 Ultra and a 128GB M5 Max both clear them on paper. How much of unified memory macOS lets the GPU use is unknown; see mac-gpu-memory-limit. The practical difference between the two is bandwidth, 614GB/s versus 1.2TB/s, and what that does to speed is unknown.
Can an M5 Ultra run a 600B-parameter model?
Arithmetic says weights alone are 600e9 x 4 / 8 / 1e9 = 300 GB at 4-bit and 600 GB at 8-bit. 300 GB is above 256GB and below 512GB; 600 GB exceeds even 512GB. KV cache for such a model is unknown because no model card is cited here. This is a hypothetical parameter count, not a specific model. Apple’s 2025 newsroom claims M3 Ultra is “capable of running large language models (LLMs) with over 600 billion parameters entirely in memory”, a vendor claim. For mixture-of-experts models see giant-moe-on-mac.
Should you wait for the M5 Ultra Mac Studio?
Apple says shipping started September 22 and the 512GB configuration arrives in late October. If you need 256GB or less, the Mac Studio specs list it and Apple’s September 22 release names only the 512GB configuration as delayed; if you need 512GB, the date is Apple’s. Other memory tiers: best-llm-for-128gb-mac, which-mac-for-local-llm. Buying advice beyond these facts depends on your models, which this page cannot know. Clustering and the NVIDIA alternative are on mac-studio-vs-dgx-spark and mac-llm-cluster.
Questions
What is the M5 Ultra LLM memory limit?
Apple lists unified memory of 96GB base, configurable to 256GB or 512GB, with 512GB as the maximum; Apple says the 512GB configuration is coming in late October. How much of it a model can use is unknown.
What is the M5 Ultra memory bandwidth?
1.2TB/s per Apple’s Mac Studio spec page, which Apple says is 50 percent more than M3 Ultra. The M5 Max is listed at 460GB/s or 614GB/s depending on the GPU configuration.
How many tokens per second does the M5 Ultra do?
Unknown. Apple’s pages give prompt-processing multiples as vendor claims, not tokens per second, and this site publishes no speed figures. See measure-tokens-per-second-mac.
Is M5 Ultra better than M3 Ultra for LLMs?
On Apple’s numbers it has higher bandwidth (1.2TB/s versus 819GB/s) and a vendor claim of up to 4x faster prompt processing in LM Studio. Real-world speed on your model is unknown.
Is an M5 Max with 128GB enough instead of an M5 Ultra with 96GB?
For the 72B example above, yes on paper: both exceed the 85.437 GB 8-bit total. The Ultra has the higher bandwidth and the larger memory ceiling; the speed effect is unknown.
Can a Mac Studio run an LLM?
Yes. Apple’s Mac Studio announcement markets the M5 Ultra for running “enormous LLMs entirely on device”. Fit depends on the model total versus installed memory; use how-much-ram-for-local-llm or mac-studio-llm.
Sources read on 2026-10-13
- Apple Mac Studio specs: chips, cores, memory options, bandwidth, 512GB maximum, blank price.
- Apple newsroom, August 2026, new Mac Studio: starting prices, 512GB timing, LM Studio prompt-processing claims (vendor), “enormous LLMs entirely on device”.
- Apple newsroom, August 2026, M5 Ultra: 1.2TB/s, 50 percent over M3 Ultra, 4.5x GPU compute claim and test basis.
- Apple newsroom, September 2026, availability: availability September 22, 512GB late October.
- Apple Support 122211: M3 Ultra 96GB to 256GB, 819GB/s.
- Apple newsroom, March 2025: M3 Ultra up to 512GB, 600B-parameter claim.
- Qwen2.5-72B-Instruct card and config.json: 72.7B parameters, layers, KV heads, hidden size.