Local AI hardware tiers explained: 12 GB to 128 GB
Memory decides what you can run. 8–12 GB covers 8B–14B chat, SDXL and Whisper. 16 GB adds gpt-oss 20B and FLUX at FP8. 24–32 GB runs 27B–35B models. 64 GB unified fits Llama 3.3 70B (tight) and 128 GB unified fits gpt-oss 120B and Qwen3.5 122B.
8–12 GB 8–12 GB GPU
The entry tier: small chat models, speech-to-text and image generation.
Runs: 8B–14B chat models at Q4, SDXL, Whisper, FLUX at Q4.
Complete: A prebuilt desktop or laptop with a 12 GB GPU. Build: 12 GB GPU, 32 GB RAM, the PSU size the card listing advises, 1 TB NVMe.
From $329.99ASRock Challenger Intel Arc B580 12GB (card only) · checked Oct 27, 2026, about 12:35 PM CT
16 GB 16 GB GPU
The sweet spot for 20B-class models and FP8 image models.
Runs: gpt-oss 20B, Mistral Small 3.2 24B (tight) and FLUX at FP8, plus everything in the 8–12 GB tier.
Complete: A prebuilt with an RTX 5060 Ti 16 GB, plus a RAM upgrade to 32 GB. Build: 16 GB GPU, 32 GB RAM, 550–650 W PSU, 2 TB NVMe for models.
From $1,599.99ABS Cyclone Aqua gaming PC (Core i7-14700F, RTX 5060 Ti 16GB, 16GB DDR5, 1TB SSD) · checked Oct 27, 2026, about 12:35 PM CT
24–32 GB GPU
Workstation cards for 30B-class models fully on the GPU.
Runs: 27B–35B models at Q4 (Qwen3.5 27B, Gemma 4 31B, DeepSeek-R1 32B).
Complete: A prebuilt AI workstation with a 32 GB card. Build: 24–32 GB GPU, 64 GB RAM, the PSU size the card listing advises, 2–4 TB NVMe.
Check price at store
Unified 64 Unified 64 GB
Mini PCs whose GPU can use most of a big shared memory pool.
Runs: 35B MoE models; Llama 3.3 70B at Q4 (tight).
Complete: A Strix Halo 64 GB mini PC, or a Mac mini with 64 GB when in stock. Build: None: unified memory is soldered. The DIY alternative is two GPUs.
From $2,599Minisforum MS-S1 MAX 64GB / 2TB (Ryzen AI Max+ 395) · checked Oct 27, 2026, 12:40 PM CT
Unified 128 Unified 128 GB
The biggest single-box tier for 100B-class open models.
Runs: gpt-oss 120B, Qwen3.5 122B-A10B, Llama 3.3 70B with room.
Complete: A Strix Halo 128 GB mini PC. Build: None: unified memory is soldered. It can join a cluster over 10 GbE or USB4.
From $3,919Minisforum MS-S1 MAX 128GB / 2TB (Ryzen AI Max+ 395) · checked Oct 27, 2026, 12:40 PM CT
How we place a model in a tier
We take the model's download size from the Ollama library at Q4_K_M (or the format it ships in), add roughly 1–4 GB for context and runtime, and pick the smallest tier that holds it. "Tight" means it should fit at short context but may spill into system RAM at longer ones. The headroom figure is our estimate.
FAQ
How much VRAM do I need to run AI locally?
8–12 GB runs 8B–14B chat models at Q4, SDXL and Whisper. 16 GB adds gpt-oss 20B. 24–32 GB runs 27B–35B models. 70B and larger models need 64–128 GB of unified memory or several GPUs. These are our tier calls built on Ollama download sizes.
Is unified memory as fast as a GPU?
Generally no. Unified-memory boxes fit bigger models but have less memory bandwidth than mid-range GPUs, so they generate tokens more slowly. We will publish numbers only from our own runs.