Local-AI beasts: boxes that run 70B+ models
Updated 2026-10-27
If you want your agent's model to run on your own desk, with no API bills and no data leaving the house, you need a lot of fast memory. Every box here has 96GB or more of unified memory. They fall into three families:
- AMD Ryzen AI Max+ 395 ("Strix Halo"), with 128GB of LPDDR5x, sold by GMKtec, Beelink, Minisforum, ACEMAGIC and Framework. Published result: 46.05 tok/s on gpt-oss-120b (MXFP4).
- NVIDIA GB10 Grace Blackwell (DGX Spark and the ASUS Ascent GX10), with 128GB and the CUDA stack.
- Apple Mac Studio M5 Ultra, which goes up to 512GB of unified memory on higher configurations.
| Box | Maker | RAM | From | Tier | Benchmark |
|---|---|---|---|---|---|
| ACEMAGIC M1A PRO+ (AI Max+ 395) AMD Ryzen AI Max+ 395 |
ACEMAGIC | 128GB | $3,399 In stock |
Big local (70B+) | 46.05 tok/s |
| Framework Desktop (Max+ 395, 128GB) AMD Ryzen AI Max+ 395 |
Framework | 128GB LPDDR5x | $3,449 Pre-order |
Big local (70B+) | 46.05 tok/s |
| GMKtec EVO-X2 (AI Max+ 395, 128GB) AMD Ryzen AI Max+ 395 |
GMKtec | 128GB LPDDR5x | $3,499.99 In stock |
Big local (70B+) | 46.05 tok/s |
| GMKtec EVO-X3 (AI Max+ 395) AMD Ryzen AI Max+ 395 |
GMKtec | 128GB LPDDR5x | $3,799.99 In stock |
Big local (70B+) | 46.05 tok/s |
| Minisforum MS-S1 MAX (128GB) AMD Ryzen AI Max+ 395 |
Minisforum | 128GB LPDDR5x (soldered) | $3,919 In stock |
Big local (70B+) | 46.05 tok/s |
| Beelink GTR9 Pro (AI Max+ 395) AMD Ryzen AI Max+ 395 (16C/32T, Radeon 8060S) |
Beelink | 128GB LPDDR5x-8000 (soldered) | $4,349 In stock |
Big local (70B+) | 46.05 tok/s |
| Mac Studio (M5 Ultra) Apple M5 Ultra, 30-core CPU / 64-core GPU |
Apple | 96GB unified (up to 512GB on 80-core GPU) | $5,499 Available |
Big local (70B+) | — |
| ASUS Ascent GX10 NVIDIA GB10 Grace Blackwell |
ASUS | 128GB LPDDR5x unified | $5,999 | Big local (70B+) | 52.87 tok/s |
| GMKtec EVO-X5 Pro (AI Max+ PRO 495) AMD Ryzen AI Max+ PRO 495 |
GMKtec | 192GB | $6,799 In stock |
Big local (70B+) | — |
| NVIDIA DGX Spark NVIDIA GB10 Grace Blackwell (20-core Arm) |
NVIDIA | 128GB LPDDR5x unified (273GB/s) | $6,950* Out of stock (per tracker) |
Big local (70B+) | 52.87 tok/s |
Benchmarks are token-generation speeds from the linked sources, on the listed machine or one with the same chip and memory. Prompt processing speed differs a lot between these families and matters for long agent contexts; see the sources.