The best Macs and mini PCs for running LLMs locally (2026)

A Mac Studio, an NVIDIA DGX Spark or a mini PC built on AMD’s Ryzen AI Max+ 395 can give its GPU 96GB or more of memory, several times the 24GB to 32GB on a consumer graphics card. That lets one quiet box run models like gpt-oss-120b and Qwen3.5 122B-A10B, which would otherwise need several GPUs or a $16,000 workstation card.

The catch is speed. These machines share LPDDR5X memory between the CPU and GPU, and it is slower than the GDDR memory on a graphics card. So they are excellent for mixture-of-experts (MoE) models, which read only a small part of their weights for each token, and slow on large dense models. If the models you want fit in 24GB to 32GB, a GPU desktop is the better buy; our GPU buying guide covers those.

The trade-off: capacity versus bandwidth

To write each token, a model reads its active weights from memory, so generation speed is roughly memory bandwidth divided by the bytes read per token. A dense model reads all of its weights every time. An MoE model reads only the experts it routes to: gpt-oss-120b has 117B parameters but only 5.1B active, so each token touches a small slice of its 65GB of weights.

That is why the same machine can feel fast or slow depending on the model. Estimated, a Ryzen AI Max+ 395 runs gpt-oss-120b at ~35 tok/s but the dense Qwen3.8 27B at Q4_K_M at only ~11 tok/s, even though the second model is a quarter of the size. The Mac Studio M5 Max has 614 GB/s, about 2.4 times the Strix Halo’s 256 GB/s, and runs gpt-oss-120b at ~64 tok/s. The M5 Ultra doubles that bandwidth again.

Generation speed

gpt-oss-120b, fully in memory

Estimated tokens per second, single user, higher is better.

1. RTX PRO 6000

   96GB · 1,792 GB/s

   ~170tok/s

2. M5 Ultra 256GB

   256GB · 1,228 GB/s

   ~93tok/s

3. M3 Ultra 512GB

   512GB · 819 GB/s

   ~70tok/s

4. M5 Max 128GB

   128GB · 614 GB/s

   ~64tok/s

5. M4 Max 128GB

   128GB · 546 GB/s

   ~58tok/s

6. DGX Spark

   128GB · 273 GB/s

   ~37tok/s

7. Ryzen AI Max+ PRO 495

   192GB · 273 GB/s

   ~37tok/s

8. Ryzen AI Max+ 395

   128GB · 256 GB/s

   ~35tok/s

Estimates from our speed model. Measured results are on the benchmarks page.

The models that suit these machines are MoE designs: gpt-oss-120b (about 68GB with 8K of context), Qwen3.5 122B-A10B (about 77GB at Q4_K_M), GLM-4.5-Air (about 68GB at Q4_K_M) and Qwen3.6 35B-A3B at high quality. With 256GB or more you can step up to DeepSeek V4 Flash (about 168GB at its native FP4) and GLM-5.3-Flash (about 161GB at Q3_K_M). Speeds in this guide are estimates from our estimate model, which assumes 8K of context and one user; we have not tested these machines ourselves.

Prompt processing is the other half

Generation speed is only half of the wait. Before a model writes anything, it has to read your prompt, and that step depends on raw compute rather than memory bandwidth. For a short chat message the difference barely registers. For a coding agent that sends a large chunk of your codebase with every request, or a long PDF, it can mean a long pause before the first word appears.

Here the machines differ more than their bandwidth suggests. Fast NVIDIA graphics cards are quicker than all of them. Among the small machines, DGX Spark is the strongest, much faster than Strix Halo. Apple says the M5 generation adds a Neural Accelerator to each GPU core, which speeds up a step where earlier Macs were notably slow. Our estimates cover generation only, so check measured prompt speeds on our benchmarks page if long inputs are your main workload.

How much memory can the GPU actually use?

None of these machines hands all of its memory to the GPU by default. The operating system needs its share, and each platform sets the limit differently.

Macs. macOS lets the GPU use about 75 percent of unified memory by default. You can raise that with the iogpu.wired_limit_mb setting (for example, sudo sysctl iogpu.wired_limit_mb=110000 on a 128GB machine), which lasts until you restart. Leave enough for macOS and your apps, or the system can slow to a crawl or become unstable.

Strix Halo. On Windows you set the GPU’s share in the BIOS or AMD’s driver software, up to 96GB. On Linux the GPU can use about 120GB after you change a kernel setting, which is enough for Qwen3.5 122B-A10B at Q6_K (about 105GB) or the 3-bit DeepSeek V4 Flash builds of about 103GB to 110GB. Our estimates assume the 96GB Windows limit.

DGX Spark. Linux reports about 119GB, and the CPU and GPU draw from the same pool.

Ryzen AI Max+ PRO 495. AMD hasn’t said how much of the 192GB the GPU can use. Our estimates assume 75 percent, about 144GB, until it does.

Software: MLX, CUDA, ROCm and Vulkan

On a Mac, llama.cpp runs through Apple’s Metal API, and MLX is Apple’s own machine learning framework, with MLX builds of most popular models published on Hugging Face. LM Studio can use either, and Ollama works too. What you can’t run is anything that needs CUDA, which rules out some fine-tuning libraries and serving tools.

DGX Spark ships with DGX OS, NVIDIA’s Ubuntu-based Linux, and the full CUDA stack, so vLLM, TensorRT-LLM, PyTorch and llama.cpp all run. Its CPU is Arm-based, so the occasional x86-only tool or container image needs a workaround.

Strix Halo machines run Windows or Linux like any PC. llama.cpp’s Vulkan backend works with standard AMD drivers and is the simplest route, including in LM Studio on Windows; ROCm is the other option on Linux. Whatever you use, keep runtimes current: newer models such as Qwen3.5 and DeepSeek V4 Flash use attention designs that need recent llama.cpp, MLX or vLLM builds.

Best overall: Mac Studio M5 Max 128GB

The Mac Studio M5 Max with the 40-core GPU and 128GB has 614 GB/s of bandwidth, more than twice what Strix Halo or DGX Spark offer, and about 96GB of it is available to the GPU by default. It’s the best balance of capacity and speed in this class, in a quiet desktop you can also use as your everyday computer.

Estimated, it runs gpt-oss-120b at ~64 tok/s, Qwen3.5 122B-A10B at Q4_K_M at ~34 tok/s and GLM-4.5-Air at Q5_K_M at ~25 tok/s. Qwen3.6 35B-A3B runs at Q8_0 at ~56 tok/s. Dense models are slower: Qwen3.8 27B at Q4_K_M runs at ~23 tok/s, about 60 percent of a used RTX 3090’s speed.

It costs about $5,100 as configured (the M5 Max Mac Studio starts at $2,499 with 36GB), the memory can’t be upgraded later, and there is no CUDA. Order the 40-core GPU, because the 614 GB/s figure applies to that version. The M4 Max 128GB it replaced has 546 GB/s, about 11 percent less, and runs gpt-oss-120b at ~58 tok/s; it launched at $3,499 in this configuration, so a refurbished or used one is a good buy if it sells well below the M5 Max. For head-to-heads, see DGX Spark vs M5 Max and Ryzen AI Max+ 395 vs M5 Max.

Best value: Ryzen AI Max+ 395 mini PCs

The Ryzen AI Max+ 395, known as Strix Halo, pairs 16 Zen 5 cores with a 40-CU Radeon 8060S GPU and up to 128GB of LPDDR5X-8000 at 256 GB/s. Several brands sell it in small desktops, and the chip runs at 45 to 120W, so these machines stay quiet. It’s also a complete x86 PC, which makes it the easiest of these machines to fit into an existing Windows or Linux setup.

Estimated, gpt-oss-120b runs at ~35 tok/s, Qwen3.5 122B-A10B at Q4_K_M at ~18 tok/s and GLM-4.5-Air at Q5_K_M at ~13 tok/s. Qwen3.6 35B-A3B at Q4_K_M runs at ~49 tok/s, which is quick enough for interactive coding help.

Dense models are where it struggles: Qwen3.8 27B at Q4_K_M runs at ~11 tok/s. Prompt processing is slower than on DGX Spark, so long inputs take a while. And the price has climbed: 128GB systems sold for about $3,300 to $4,350 in October 2026, up from about $2,000 in 2025. It is still the cheapest new machine that runs 100B-class models entirely in GPU-accessible memory.

Best for CUDA developers: NVIDIA DGX Spark

DGX Spark is NVIDIA’s compact desktop built on the GB10 Grace Blackwell chip, with a 20-core Arm CPU, a Blackwell GPU and 128GB of LPDDR5X at 273 GB/s. Linux reports about 119GB usable, more than a 128GB Mac gives its GPU by default. Two units can be linked through the built-in ConnectX-7 networking.

Generation speed is in the same class as Strix Halo. Estimated, gpt-oss-120b runs at ~37 tok/s, Qwen3.5 122B-A10B at Q4_K_M at ~19 tok/s (or at Q6_K, which fits in the extra memory, at ~14 tok/s), and Qwen3.6 35B-A3B at Q8_0 at ~32 tok/s. Where the Spark pulls ahead is prompt processing, which is much faster than on Strix Halo, and software. It runs the same CUDA stack as NVIDIA’s datacenter systems, so work you prototype on it carries over to servers.

The problem is price. It launched at $3,999 in October 2025, rose to $4,699 in February 2026 and listed at about $6,950 in October 2026, more than twice a Strix Halo mini PC for similar generation speed. A 64GB OEM version is due from $4,999 on October 23, but 64GB is too small for gpt-oss-120b, which needs about 68GB. Buy the Spark if CUDA matters to your work; otherwise, see DGX Spark vs Ryzen AI Max+ 395.

Best for the biggest models: Mac Studio M5 Ultra 256GB

The Mac Studio M5 Ultra with 256GB has about 1.2TB/s of bandwidth, by Apple’s figure, and about 192GB of GPU memory by default. At $9,499 it costs nearly twice the M5 Max, but it doubles both the memory and the bandwidth, and that opens up a class of models that 128GB machines can’t hold at full quality.

Estimated, it runs DeepSeek V4 Flash at its native FP4 (about 168GB) at ~45 tok/s and GLM-5.3-Flash at Q3_K_M (about 161GB) at ~39 tok/s. Smaller models get fast: gpt-oss-120b at ~93 tok/s and Qwen3.5 122B-A10B at Q8_0 at ~33 tok/s. Even dense Qwen3.8 27B at Q4_K_M reaches ~39 tok/s, the same as a used RTX 3090. For comparison, a pair of RTX PRO 6000 cards with 192GB runs DeepSeek V4 Flash at ~72 tok/s, for about $32,000 in cards alone.

The 512GB M5 Ultra ships in late October, and Apple hadn’t announced its price as of mid-October. It is the only current desktop that holds GLM-5.3, at Q3_K_M in about 375GB, which only just fits under the default GPU limit (~19 tok/s, estimated), or GLM-5.3-Flash at Q6_K (~25 tok/s). If you don’t need that much, the M5 Ultra also starts at $5,499 with 96GB. A used M3 Ultra 512GB has 819 GB/s and runs gpt-oss-120b at ~70 tok/s, but used prices are volatile, so check completed sales first.

How they compare

GPU	120B MoE speed	122B MoE speed	27B dense speed	Price
M5 Max 128GB128GB unified · 614 GB/s	~64tok/s	~34tok/s	~23tok/s	~$5,100Configure at Apple ↗
Ryzen AI Max+ 395128GB unified · 256 GB/s	~35tok/s	~18tok/s	~11tok/s	~$3,300Check Amazon ↗
DGX Spark128GB unified · 273 GB/s	~37tok/s	~19tok/s	~12tok/s	~$6,950View at NVIDIA ↗
M5 Ultra 256GB256GB unified · 1,228 GB/s	~93tok/s	~53tok/s	~39tok/s	~$9,499Configure at Apple ↗
M5 Ultra 512GB512GB unified · 1,228 GB/s	~93tok/s	~53tok/s	~39tok/s	—Configure at Apple ↗
Ryzen AI Max+ PRO 495192GB unified · 273 GB/s	~37tok/s	~19tok/s	~12tok/s	~$6,800Check Amazon ↗
RTX 309024GB · 936 GB/s	~21tok/s	Won't fit	~39tok/s	~$1,050 usedCheck Amazon ↗
RTX 509032GB · 1,792 GB/s	~25tok/s	~11tok/s	~80tok/s	~$4,300Check Amazon ↗

Also considered: Ryzen AI Max+ PRO 495

The Ryzen AI Max+ PRO 495, known as Gorgon Halo, raises memory to 192GB but keeps the same 256-bit bus, so bandwidth only rises to 273 GB/s. It runs the same models as Strix Halo at about the same speed (Qwen3.5 122B-A10B at Q6_K at ~14 tok/s, estimated). The extra memory buys higher-quality quantizations, or room for 200B-class models at around 3 bit, but nothing in speed. At $6,799 for the Framework Desktop (shipping in November) or $7,399 for the Minisforum MS-S1 MAX-P495, the $9,499 M5 Ultra 256GB costs 40 percent more and has about 4.5 times the bandwidth. We’d wait until AMD documents how much of the memory the GPU can use.

When a GPU desktop is the better buy

Unified memory wins when you want 100B-class MoE models, low noise and one simple box. A GPU desktop wins almost everywhere else.

Generation speed

Qwen3.8 27B at Q4_K_M: a dense model that fits on one GPU

Estimated tokens per second, single user, higher is better.

1. RTX 5090

   32GB · 1,792 GB/s

   ~80tok/s

2. RTX 3090

   24GB · 936 GB/s

   ~39tok/s

3. M5 Ultra 256GB

   256GB · 1,228 GB/s

   ~39tok/s

4. M5 Max 128GB

   128GB · 614 GB/s

   ~23tok/s

5. DGX Spark

   128GB · 273 GB/s

   ~12tok/s

6. Ryzen AI Max+ 395

   128GB · 256 GB/s

   ~11tok/s

Estimates from our speed model. Measured results are on the benchmarks page.

If your models fit in 24GB to 32GB. The strongest models that fit on one card, such as Qwen3.8 27B and Gemma 4 31B, are dense. A used RTX 3090 at about $1,050 runs Qwen3.8 27B at ~39 tok/s (estimated), against ~23 tok/s on the M5 Max and ~11 tok/s on Strix Halo. An RTX 5090 reaches ~80 tok/s.

If you work with long prompts. NVIDIA GPUs read prompts far faster, which matters for coding agents and document work.

If you need CUDA or plan to fine-tune. Training tools and many serving stacks assume an NVIDIA GPU. DGX Spark covers CUDA, but for models that fit in a graphics card’s memory, a desktop GPU is faster for less money.

If you want to upgrade later. You can add a second card to a desktop. Unified memory is fixed the day you buy it.

Use the VRAM calculator to check what your target models need, and our guide to how much VRAM you need to decide how much headroom to buy.