Best hardware for local AI (2026)
Not sure? Our local AI PC picks by budget, from €800 to €3,500 →
Which machine should you buy to run AI models locally ? We compare the serious options for 2026 — compact AI workstation, 128 GB mini-PC, and dedicated GPU — based on what really matters: the memory, its bandwidth, and price. Independent buying verdict.
Updated on 09/10/2026
Comparison at a glance
| Hardware | Memory | Bandwidth | Indicative price | Type |
|---|---|---|---|---|
| NVIDIA DGX Spark | 128 GB unified (LPDDR5x) | 273 GB/s | ≈ 6,100 € at Materiel.net (DGX Spark PNY at 6,099.95 € on 28/09/2026, lead time of more than 15 days; the MSI EdgeXpert, with the same GB10 chip and 128 GB, is 6,599.95 € with delivery within 7 days) | Compact AI workstation |
| GIGABYTE AI TOP ATOM | 128 GB unified (LPDDR5x), 121.7 GB visible to the system | 273 GB/s | 6 346,27 € (ATAGB10-9001, 4 TB PCIe 4.0, official AORUS store, sold out; in stock on Amazon.fr) · 7 999,95 € (ATAGB10-9000, 4 TB PCIe 5.0, LDLC and Materiel.net, out of stock) — recorded on 08/10/2026 | Compact AI workstation |
| NVIDIA RTX Spark | 128 GB of unified CPU/GPU-shared memory (GB10) | LPDDR5x unified memory—SoC bandwidth, lower than the GDDR7 in a RTX 5090 (manufacturer figure not provided for laptops) | ≈ €2,660 announced for the first configurations (press, fall 2026 — to be confirmed) | AI SoC for laptops and compact desktops |
| Ryzen AI Max+ 395 Mini PC (Strix Halo) | Up to 128 GB unified memory (LPDDR5X-8000) | ≈ 212 GB/s | ≈ 2 000 € for 64 GB and ≈ 3 300 € for 128 GB (GMKtec EVO-X2, official European store, September 2026) | 128 GB AI mini PC / workstation |
| Mac Studio M4 Max | 36 GB (32-core GPU) to 96 GB unified (128 GB+ tiers remain removed) | 410 GB/s (32-core GPU) · 546 GB/s (40-core GPU) | €2 219 at the latest check (36 GB / 512 GB, Darty), sold out by the end of September 2026 | Apple workstation (unified memory) |
| Mac mini M6 | 16, 24, or 32 GB unified memory (24 GB recommended for local AI) | 170 GB/s (+10% vs. M5) | €1,049 (16 GB / 256 GB) · €1,489 (24 GB / 512 GB, the local AI configuration) · +€440 for 32 GB | Apple mini workstation (unified memory) |
| Mac Studio M5 Max / M5 Ultra | M5 Max: 36 to 128 GB · M5 Ultra: 96 to 512 GB unified (the 512 GB tier arrives in late October) | M5 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) | €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 | Apple workstation (unified memory) |
| Mac mini M5 Pro | 24 to 64 GB unified (LPDDR5X) | 307 GB/s | €1,999 (24 GB / 512 GB); 48–64 GB configurations in BTO Apple | Apple mini workstation (unified memory) |
| MacBook Pro M5 Max | 36 GB (32-core GPU) to 128 GB unified (40-core GPU) | 460 GB/s (32-core GPU) · 614 GB/s (40-core GPU) | starting at €4,199 (36 GB / 32-core GPU); 48–128 GB with a 40-core GPU (BTO Apple) | Portable Apple (unified memory) |
| AMD Radeon RX 9070 XT 16 GB | 16 GB GDDR6 | 644 GB/s | ≈ 860 – 1 070 € | Dedicated graphics card |
| RTX laptop (5080/5090) for local AI | 16 GB (RTX 5080) or 24 GB (RTX 5090) dedicated GDDR7 | ≈ 700–900 GB/s depending on the chassis TGP | from ≈ €2,500 (RTX 5080 16 GB); from ≈ €4,300 (RTX 5090 24 GB) | Gaming/AI laptop (dedicated GPU) |
How do you choose?
NVIDIA DGX Spark
The compact AI workstation from NVIDIA, with 128 GB of unified memory.
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.
GIGABYTE AI TOP ATOM
GIGABYTE's AI mini PC: NVIDIA GB10 chip and 128 GB of unified memory in one liter.
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.
NVIDIA RTX Spark
The DGX Spark chip in a laptop: 128 GB unified memory, CUDA, fall 2026.
Buying alternative for local AI. The button configuration is the one offered for purchase. On mini PCs, reserve memory for the system and check the inference engine; macOS/MLX and CUDA are not interchangeable.
Ryzen AI Max+ 395 Mini PC (Strix Halo)
The best memory-per-euro value for local AI.
Another available configuration:
AmazonGMKtec EVO-X2 64 GB / 1 TB (Ryzen AI Max+ 395) →Choose the capacity shown on the button based on the memory your models need. This memory is shared with the system and selected at purchase: it cannot be added later.
GMKtec: official European store (French-language site), with delivery in France averaging 5 to 8 days.
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.
Read the full guide: Ryzen AI Max+ 395 mini-PC (Strix Halo) →
Mac Studio M4 Max
The previous-generation Apple workstation for local AI—replaced by the Mac Studio M5.
Buying alternative for local AI. The button configuration is the one offered for purchase. On mini PCs, reserve memory for the system and check the inference engine; macOS/MLX and CUDA are not interchangeable.
Mac mini M6
The first 2 nm Mac — and the cheapest entry point to local AI from Apple.
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.
Mac Studio M5 Max / M5 Ultra
The first consumer machine to run a frontier model locally.
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.
Mac mini M5 Pro
The most accessible Mac for local AI — now with the M5 Pro.
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.
MacBook Pro M5 Max
The Apple laptop supports up to 128 GB for local AI.
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.
AMD Radeon RX 9070 XT 16 GB
A dedicated 16 GB GPU with a good price-to-VRAM ratio.
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.
RTX laptop (5080/5090) for local AI
Local AI that fits in a bag—16 or 24 GB of VRAM.
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.
Read the complete guide: RTX laptop (5080/5090) for local AI →
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.
Frequently asked questions
What hardware do you need to run a 70B LLM locally?
You need either ~48 GB of VRAM (multiple GPUs) or a machine with 128 GB unified memory : Mac Studio, Ryzen AI Max+ 395 mini-PC (Strix Halo), or NVIDIA DGX Spark. Memory capacity and bandwidth are factor n°1.
Mac, PC with a dedicated GPU, or DGX Spark–style AI workstation?
A dedicated GPU (e.g., RX 9070 XT 16 GB) is fast but limited to models ≤ 16 GB. 128 GB of unified memory (Mac / Strix Halo / DGX Spark) unlocks large models. The DGX Spark is primarily aimed at the CUDA / on-premises ecosystem; the Strix Halo mini PC offers the best memory-per-euro ratio.
How much memory for local AI?
8–16 GB for small models (7–14B), 24–32 GB for a comfortable 32B, 64–128 GB for targeting a 70B or large MoE models.