Mac Studio M5 Max / M5 Ultra: local AI test and review
The first consumer machine capable of running a frontier model locally. We look at what matters for running LLMs locally : memory, bandwidth, speed, price—and who it's (or isn't) the right purchase for.
Updated on 09/10/2026
Where to buy it
The button configuration is the one offered for purchase. On mini PCs, keep some memory available for the system and check the inference engine; macOS/MLX and CUDA are not interchangeable.
Affiliate links — QuelLLM may earn a commission on purchases, at no extra cost to you, which does not influence these recommendations (established independently). As an Amazon Associate, BestLLMfor earns from qualifying purchases.
Technical specifications
- MemoryM5 Max: 36 to 128 GB · M5 Ultra: 96 to 512 GB unified (the 512 GB tier arrives in late October)
- BandwidthM5 Max: 460 GB/s (32-core GPU) / 614 GB/s (40-core GPU) · M5 Ultra: up to 1.2 TB/s (+50% vs. previous generation)
- ComputeM5 Max: 18-core CPU, up to 40-core GPU · M5 Ultra: 36-core CPU, 80-core GPU, 32-core Neural Engine — Neural Accelerators in every GPU core, Thunderbolt 5, PCIe Gen 6 SSD
- Power consumptionDesktop workstation, quiet cooling
- Indicative price€2,999 (M5 Max 36 GB / 512 GB) · €6,599 (M5 Ultra 96 GB / 1 TB) — up to ≈ €20,000 with 512 GB / 16 TB
- Largest model36 GB: 30B class · 64-96 GB: 70B in Q4 · 128 GB: GPT-OSS-120B comfortably · 256 GB: Qwen3-235B (MoE) · 512 GB: ~600B frontier MoE in Q4, DeepSeek class — see tier table
Who is it for?
Advantages and limitations
✓ Strengths
- 512 GB of unified memory at 1.2 TB/s: a ~600-billion-parameter model (Q4) fits ENTIRELY locally — unprecedented outside a server
- Apple claims up to 4.3× the AI performance of the previous generation (Neural Accelerators per GPU core)
- macOS + MLX: the most optimized local LLM ecosystem for Apple (LM Studio, Ollama, mlx-lm)
- Quiet, compact, and low-power for its capacity — no multi-GPU tower can house 512 GB of VRAM
- Thunderbolt 5 and PCIe Gen 6 SSD: significantly faster weight loading
✗ Limitations
- The 512 GB tier does not ship until late October 2026 (the rest from 09/22)
- Price Apple: the 512 GB configuration far exceeds €10,000—reserved for serious professional use
- Below 96 GB, a 128 GB Strix Halo mini PC (≈ €3,300 to €4,000) offers more memory per euro
- No CUDA: heavy fine-tuning and some tools remain simpler on the NVIDIA side
Which memory tier for which models?
The rule: the memory sets the size of the model that can be loaded, the bandwidth determines generation speed. Speeds are given as rough orders of magnitude for text generation (decoding), Q4 quantization, via MLX / llama.cpp — MoE models use only their active parameters, hence their high throughput.
| Configuration | Bandwidth | What runs (Q4) | Estimated speed |
|---|---|---|---|
| M5 Max 36 GB | 460 GB/s | 30B class: Qwen3-32B, GLM-4.7-Flash | ~25–35 tok/s |
| M5 Max 64 GB | 614 GB/s (40-core GPU) | 70B in Q4 (~40 GB); very fast 30B-A3B MoE | ~13–16 tok/s (70B) · 80+ tok/s (MoE A3B) |
| M5 Max 128 GB | 614 GB/s | 70B in Q8, GPT-OSS-120B (MXFP4), long context | ~40-60 tok/s (GPT-OSS-120B) |
| M5 Ultra 96 GB | 1.2 TB/s | 70B in Q4 with ample context headroom | ~25-30 tok/s |
| M5 Ultra 256 GB | 1.2 TB/s | Qwen3-235B-A22B in Q4 (~135 GB) | ~25–35 tok/s |
| M5 Ultra 512 GB | 1.2 TB/s | ~600B frontier MoE in Q4 (class DeepSeek, ~380-400 GB) | ~15-20 tok/s |
Prefill (reading the prompt) also benefits from the M5 GPU's Neural Accelerators—this was historically the weak point of Macs compared with NVIDIA GPUs on long contexts.
M5 Max or M5 Ultra?
Le M5 Max (€2,999 for 36 GB, BTO up to 128 GB) covers everything up to 70B and ~120B MoE models: it is the rational choice for demanding dev/copilot/RAG use. Be sure to get the 40-core GPU as soon as you exceed 36 GB: bandwidth increases from 460 to 614 GB/s, delivering one-third more speed for each token.
Le M5 Ultra (€6,599 in 96 GB) is justified only by its memory: at 96 GB it is barely better than a cheaper 128 GB M5 Max. Its real purpose is the tiers 256 and 512 GB — the only ones to run MoE models from 235 to ~600 billion parameters. If you don't have a specific use for these models, the Ultra is a bad purchase; if you do, it's the only desktop hardware that can run them.
Compared with the alternatives
| Machine | Memory | Bandwidth | Pricing | Using it properly |
|---|---|---|---|---|
| Mac Studio M5 Max | 36-128 GB | 460–614 GB/s | starting at €2,999 | Fast 70B and ~120B MoE models, MLX ecosystem |
| Mac Studio M5 Ultra | 96-512 GB | 1.2 TB/s | from €6,599 | MoE 235–600B: unmatched on desktop |
| Strix Halo Mini-PC | 64–128 GB | 212 GB/s | ≈ 2 000-4 000 € | Best memory-to-euro ratio (70B, slow but accessible) |
| NVIDIA DGX Spark | 128 GB | 273 GB/s | ≈ 6 100 – 6 600 € | CUDA ecosystem required, fine-tuning |
| RTX 5090 | 32 GB | 1.79 TB/s | ≈ 5 700 € | Raw speed up to ~32B, but no further |
In summary: pure speed on small models → dedicated GPU; maximum capacity per dollar → Strix Halo; maximum capacity overall → M5 Ultra.
Availability: what you need to know
- On sale since September 22, 2026 (preorders opened on August 25).
- The tier 512 GB won't ship until the end of October — order early if that's your target.
- At French retailers: M5 Max 36 GB / 512 GB at Darty and M5 Max 96 GB / 512 GB at Materiel.net; the other memory tiers are available through the Apple Store's BTO configurator.
- The Mac Studio M4 Max 36 GB / 512 GB is no longer in stock new from our retailers (verified in late September 2026) — our M4 Max sheet to compare.
Verdict
Frequently asked questions
Can the Mac Studio M5 Max / M5 Ultra run a 70B LLM locally?
36 GB: 30B class · 64–96 GB: 70B in Q4 · 128 GB: GPT-OSS-120B with room to spare · 256 GB: Qwen3-235B (MoE) · 512 GB: frontier MoE ~600B in Q4, DeepSeek class — see the tier table.
How much does the Mac Studio M5 Max / M5 Ultra cost?
Expect €2,999 (M5 Max 36 GB / 512 GB) · €6,599 (M5 Ultra 96 GB / 1 TB) — up to ≈ €20,000 in 512 GB / 16 TB (starting at $2,499 (M5 Max) · $5,499 (M5 Ultra)). Prices change quickly — check the current price through the purchase links.
Who is the Mac Studio M5 Max / M5 Ultra for?
MacOS developers and creators targeting large LLMs on a fixed workstation—from the 30B class (36 GB) up to ~600B frontier MoE models in Q4 (512 GB).
Can you run a DeepSeek-class model locally on the Mac Studio M5?
Yes, this is THE breakthrough: with 512 GB of unified memory, a frontier MoE model with ~600 billion parameters quantized in Q4 (~380-400 GB of weights) fits entirely locally, with an estimated generation speed of 15-20 tok/s thanks to the MoE architecture (only the active parameters are read for each token). See also our DeepSeek V4 Flash guide.
Should you still choose the Mac Studio M4 Max over the M5?
No longer available new from our retailers: M4 Max 36 GB / 512 GB, cleared out at €2,219 over the summer, is no longer in stock at the end of September 2026. The M5 Max takes over at €2,999: higher bandwidth, Neural Accelerators, and memory tiers up to 128 GB that the M4 Max no longer offers.
Why does unified memory beat a GPU for large models?
A dedicated GPU tops out at 32 GB of VRAM (RTX 5090): beyond that, it has to offload to RAM over the PCIe bus, and speed collapses. Unified memory Apple gives the GPU direct access to the full 96-512 GB: the entire model stays at full bandwidth. It's slower than GDDR7 VRAM of the same size, but it's the only desktop approach that can “fit” giant models.