How to compare these machines
As covered on the local AI hub, memory capacity decides which models fit (about 0.5 to 0.6 GB per billion parameters at 4-bit, plus context) and memory bandwidth sets the ceiling on how fast tokens come out. Compute matters more for prompt processing, batching and fine-tuning. So read the table below left to right: capacity, then bandwidth, then price.
The comparison
| Machine | Memory | Memory bandwidth | Price as listed (20 Oct 2026) | What fits |
|---|---|---|---|---|
| Mac Studio, M5 Max | 36–128 GB unified | 460 GB/s (614 GB/s with 40-core GPU) | From $2,499 (apple.com) | Up to ~100–120B params at 4-bit with 128 GB; gpt-oss-120b class |
| Mac Studio, M5 Ultra | 96–512 GB unified (512 GB "coming late October") | 1.2 TB/s | From $5,499 (96 GB); $6,799 (36/80-core); top listed config $18,299; 512 GB not yet priced | With 512 GB: several hundred billion params at 4-bit, e.g. GLM-5.3-Flash (320B) with room for context |
| NVIDIA DGX Spark | 128 GB LPDDR5x (64 GB OEM version from 23 Oct) | 273 GB/s | Founders Edition MSRP $4,699 since Feb 2026 (NVIDIA); $6,950 for 128 GB reported Oct 2026 | NVIDIA: up to 200B params; two linked: up to 400B |
| NVIDIA DGX Station (GB300) | 748 GB coherent (252 GB HBM3e + 496 GB LPDDR5X) | 7.1 TB/s (GPU memory) | Not listed; through OEM partners | Trillion-parameter class at low precision; a server on a desk |
| NVIDIA RTX PRO 6000 Blackwell | 96 GB GDDR7 per card | 1,792 GB/s | Not listed by NVIDIA; ~$8,500 at launch, >$10,000 street in 2026 (reported) | ~120B MoE at 4-bit per card; more with multiple cards |
| AMD Ryzen AI Max+ 395 (Framework Desktop) | 64 or 128 GB LPDDR5X; 192 GB on PRO 495 | ~256 GB/s (256-bit bus) | 128 GB board $3,449; PRO 495 from $6,799 pre-order (frame.work) | Up to ~100–120B params at 4-bit with 128 GB |
Spec sources: Apple Mac Studio specs and Apple Store; NVIDIA DGX Spark, DGX Station and RTX PRO 6000; Framework Desktop. "What fits" is my estimate from the memory rule of thumb, weights at 4-bit, not a benchmark.
Mac Studio with M5 Ultra
Apple announced the M5 Max and M5 Ultra Mac Studio on 25 August 2026. Apple's spec page lists the M5 Ultra with a 30-core CPU (configurable to 36), a 64-core GPU (configurable to 80), 1.2 TB/s of memory bandwidth and up to 512 GB of unified memory. On the US Apple Store on 20 October 2026 it started at $5,499 with 96 GB and 1 TB, and $6,799 for the 36-core CPU and 80-core GPU version. The store's top listed configuration was $18,299.
The press figures you will see for this machine conflict, and here is why. The 512 GB memory option was marked on the store as "coming late October" and could not yet be configured or priced. Numbers of around $18,000 for a 512 GB machine come from secondary sources estimating that configuration; Apple has not listed it. Re-check the store before you budget.
For local LLMs this is the machine that holds the biggest models on one desk, at the highest bandwidth of any unified-memory system here. The trade-offs are Apple's software stack (MLX and llama.cpp's Metal backend rather than CUDA) and the fact that you cannot add a GPU later.
M5 Ultra vs M3 Ultra: what the first benchmarks show
The most complete public test so far is NetworkChuck's review, which ran an Apple-loaned M5 Ultra (80-core GPU, 256 GB, 8 TB, roughly $14,000 to $15,000 as configured according to the video) headless as a local AI server against an M3 Ultra (80-core GPU, 512 GB). The figures below are his, not ours; we have not reproduced them.
| Workload (per the video) | M5 Ultra | M3 Ultra | Speed-up |
|---|---|---|---|
| Qwen3.8 27B 4-bit, 32K context: time to first token | 21.66 s | 82.85 s | 3.8x |
| Qwen3.8 27B 4-bit: generation | 43.3 tok/s | 28.1 tok/s | 1.5x |
| Qwen 14B 4-bit: time to first token | 14 s | 57 s | 4.0x |
| Whisper large-v3, 2h05m recording | 89.4 s | — | 1.7x |
| Whisper turbo, 2h05m recording | 24.4 s | — | 2.3x |
| Qwen3-VL 32B 4-bit, video frames at 1 fps | 22 s | — | 2.6x |
| FLUX.2 klein 4B, one image | 2.6 s | 10.2 s | 4.0x |
| LTX-2.5 22B video generation | — | — | 1.7x to 2.5x |
The pattern is the useful part, because it follows from the hardware. Prompt processing (the wait before the first token) multiplies matrices across every input token at once, so it is compute-bound, and the Neural Accelerators Apple added to each M5 GPU core show up as a roughly 4x gain. Token generation produces one token at a time and has to stream the full set of weights from memory for each one, so it is bandwidth-bound: the 1.5x gain tracks the move from 819 GB/s to 1.2 TB/s almost exactly. In practice that means the M5 Ultra fixes the Mac's old weak spot, long prompts and big contexts in agent and RAG workloads, while raw generation speed improves only in line with memory bandwidth.
The same video shows what a box like this is for beyond chat: batch transcription, tagging video with a vision model, and LM Studio driving DaVinci Resolve over MCP with a local 27B model. Those are workloads where an API bill grows with volume and a fixed-cost machine does not.
Is there a Mac Pro M5?
No, and a lot of people are searching for one. Apple discontinued the Mac Pro in March 2026. The last model used the M2 Ultra; apple.com/mac-pro now redirects to the general Mac page. The Mac Studio with M5 Ultra is Apple's top desktop.
NVIDIA DGX Spark
A small box built on NVIDIA's GB10 Grace Blackwell chip with 128 GB of unified LPDDR5x at 273 GB/s, up to 1 PFLOP of FP4 compute, and a ConnectX-7 port at 200 Gb/s that links two to four units. NVIDIA says one 128 GB Spark handles models up to 200 billion parameters, two up to 400 billion.
Pricing has moved more than the hardware. The Founders Edition launched at $3,999; NVIDIA raised it to $4,699 in February 2026, citing memory supply. In October The Register reported a new 64 GB version sold only through OEM partners at about $4,999, and a further increase of the 128 GB model to $6,950. I could not load NVIDIA's marketplace to confirm the current price, so treat $4,699 as the last NVIDIA-confirmed figure and $6,950 as reported.
The case for the Spark is software. If your production inference runs on NVIDIA GPUs, developing against the same CUDA stack on your desk saves a class of surprises.
AMD Strix Halo (Ryzen AI Max+ 395)
AMD's Ryzen AI Max processors pair Zen 5 CPU cores with a large integrated Radeon GPU and up to 128 GB of LPDDR5X on a 256-bit bus, about 256 GB/s. That is the same class of memory as the Spark for a lower price. Framework listed its Desktop mainboard with the Ryzen AI Max+ 395 and 128 GB at $3,449 on 20 October 2026, and a new Ryzen AI Max+ PRO 495 with up to 192 GB for pre-order from $6,799.
The Register tested an HP Z2 Mini G1a (Strix Halo) against a DGX Spark and found single-user token generation roughly tied, with the Spark much faster at time to first token, batched inference and fine-tuning, and ahead on software maturity. That matches what the specs predict: similar bandwidth, more compute on the NVIDIA side.
NVIDIA RTX PRO 6000 Blackwell
A workstation card with 96 GB of GDDR7 at 1,792 GB/s, drawing up to 600 W. It holds less than the 128 GB boxes but reads its memory six to seven times faster, so for models that fit it is the quickest option here, and two or four cards in a workstation scale capacity. NVIDIA does not list a price on the product page. Launch pricing was reported around $8,500 and 2026 street prices have been reported well above $10,000, so get a current quote.
DGX Station and tinybox: the next step up
NVIDIA's DGX Station puts a GB300 Grace Blackwell Ultra chip in a deskside tower with 748 GB of coherent memory, 252 GB of it HBM3e at 7.1 TB/s. NVIDIA does not list a price; it is sold through partners such as Dell, HP and Supermicro. For teams that want a multi-GPU box built for them, tiny corp's tinybox is now configured to order with up to four GPUs; the price depends on the configuration, so there is no single figure to quote.
What I would buy
- One person, largest possible models, quiet desk: Mac Studio M5 Ultra, with as much memory as the budget allows.
- An engineer building for NVIDIA production: DGX Spark.
- Best capacity per dollar at 128 GB: a Strix Halo machine such as the Framework Desktop.
- Fastest inference on models up to ~120B, or fine-tuning: a workstation with one or two RTX PRO 6000 cards.
- Several concurrent users in production: none of these. Use data-center GPUs with vLLM, or rent; see neoclouds compared.
Whatever you buy, pick the model first. The best local LLMs by hardware tier page lists what runs well at each memory size.
Related
- Local AI for engineering teams — when running models yourself makes sense
- Best local LLMs — which models to run on this hardware
- Ollama vs LM Studio vs llama.cpp vs vLLM — the software that runs them
- Neoclouds compared — renting GPU hours instead
Frequently asked questions
Is there a Mac Pro with M5?
No. Apple discontinued the Mac Pro in March 2026; the last model used the M2 Ultra chip from 2023. The Mac Pro page on apple.com now redirects to the general Mac page. If you searched for "Mac Pro M5", the machine you want is the Mac Studio with M5 Ultra, which Apple announced on 25 August 2026.
How much does the M5 Ultra Mac Studio cost?
On apple.com on 20 October 2026 the M5 Ultra Mac Studio started at $5,499 (30-core CPU, 64-core GPU, 96 GB memory, 1 TB storage), and $6,799 for the 36-core CPU and 80-core GPU version. The most expensive configuration on Apple's US store was $18,299. The 512 GB memory option was marked "coming late October" and had no listed price, so figures you may see for a fully loaded 512 GB machine are press estimates.
DGX Spark vs Mac Studio: which is better for local LLMs?
For running the largest models for one person, the Mac Studio: an M5 Ultra offers up to 512 GB of memory at 1.2 TB/s, against 128 GB at 273 GB/s on the Spark, so it holds bigger models and generates tokens faster. For developing software that will run on NVIDIA GPUs in production, the DGX Spark: it runs NVIDIA's CUDA stack and DGX OS, links to a second Spark over 200 Gb/s networking, and is much faster at prompt processing and fine-tuning than its bandwidth suggests.
What is AMD Strix Halo?
Strix Halo is the codename for AMD's Ryzen AI Max processors, which combine Zen 5 CPU cores and a large Radeon GPU sharing up to 128 GB of LPDDR5X memory on a 256-bit bus (about 256 GB/s). The top consumer part is the Ryzen AI Max+ 395. A newer Ryzen AI Max+ PRO 495 supports 192 GB. Systems using it include the Framework Desktop and HP Z2 Mini G1a.
How much memory do I need to run a 70B model locally?
At 4-bit quantization the weights of a 70-billion-parameter dense model take roughly 35 to 42 GB, plus several gigabytes for context. A 64 GB machine can run it with a modest context; 96 to 128 GB gives comfortable headroom. At 8-bit, plan for about 75 GB plus context.
Is the RTX PRO 6000 worth it for local AI?
It has the most bandwidth of any single card here: 96 GB of GDDR7 at 1,792 GB/s. A 96 GB card fits models like gpt-oss-120b, which OpenAI says runs on a single 80 GB GPU, and runs them much faster than the unified-memory boxes. NVIDIA does not list a price on its product page; launch pricing was reported around $8,500 and 2026 street prices have been reported well above $10,000. It also needs a workstation that can supply 600 W.
Sources: apple.com (specs and US store, retrieved 20 October 2026); nvidia.com product pages; NVIDIA developer forums (February 2026 price change); The Register; frame.work (retrieved 20 October 2026); tinygrad.org; Macworld. Prices change frequently, especially for memory-heavy systems, and are not quotes.