Model guides
Start from the model you want to run. Each guide gives the memory it needs, the tier that runs it, and new-retail picks.
What runs gpt-oss 20B? PCs and builds that fit
gpt-oss 20B is a 14 GB download in Ollama. See the 16 GB and 24 GB GPUs, prebuilts and parts that fit it, with prices checked Oct 27, 2026.
What runs gpt-oss 120B? 128 GB unified-memory mini PCs
gpt-oss 120B is a 65 GB download in Ollama, so it needs a 128 GB unified-memory box. Compare the Strix Halo mini PCs that fit it.
What runs Llama 3.3 70B? Unified-memory mini PCs vs two GPUs
Llama 3.3 70B is a 43 GB download at Q4_K_M. Compare 64 GB and 128 GB unified-memory mini PCs with a two-GPU build.
What runs Qwen3.5? Hardware for 27B, 35B-A3B and 122B-A10B
Qwen3.5 27B needs 17–20 GB, 35B-A3B needs 22 GB and 122B-A10B needs 81 GB in Ollama. Here is the hardware tier for each size.
What runs Gemma 4 31B? 24–32 GB GPUs and mini PCs
Gemma 4 31B is a 19–20 GB download in Ollama with a 256K context window. See the GPUs and systems that hold it.
What runs DeepSeek-R1 32B? GPUs and systems that fit
The DeepSeek-R1 32B distill is a 20 GB download in Ollama. See the 24–32 GB GPUs and unified-memory boxes that run it.
What runs Llama 3.1 8B, Phi-4 and Qwen3 14B? Entry GPUs
Llama 3.1 8B (4.9 GB), Phi-4 (9.1 GB) and Qwen3 14B (9.3 GB) fit 12 GB GPUs. See the entry cards and systems that run them.
Run FLUX and SDXL at home: which GPU do you need?
SDXL runs on 8 GB of VRAM per Stability AI; FLUX.1 dev needs about 12 GB at FP8 (third-party estimate). Pick the GPU tier for image generation.
Whisper on your own PC: hardware for transcription
OpenAI lists about 10 GB of VRAM for Whisper large and about 6 GB for turbo. See the GPUs that run local transcription.
10 GbE for local AI: link two boxes and a NAS for models
Networking for local AI only: 10 GbE switches and cards, USB4 links, a NAS for model files, and the llama.cpp RPC backend as a labeled experiment.