Guides
In-depth buying guides and explainers for running AI locally — which GPU runs which model, what to buy at each budget, and where the ecosystem stands right now.
-
DGX Spark value ladder: 1 box, 2 boxes, 4 boxes vs RTX Pro 6000, H100, and the new DGX Station GB300
Nine months after launch, the DGX Spark's 273 GB/s bandwidth looks less like a flaw and more like the enabler of a desk-sized cluster that runs frontier models for a fraction of the price.
-
Intel Arc Pro B70 32GB: the buyer's guide
The cheapest 32GB discrete GPU, measured: Qwen 3.6 27B at 14 tok/s on Vulkan and 22 on SYCL, what the extra VRAM unlocks on 32B models, and whether to buy at ~$949-$1,099.
-
State of Local AI
The quarterly snapshot: what's winning at each budget, dense vs MoE, what's coming, and what to actually buy right now.
-
Apple M5 Max for Local LLMs
What the 128GB M5 Max actually runs — Qwen 3.5-122B-A10B, 70B dense, and a 235B MoE no other laptop holds — the truth behind Apple's 4x prefill claim, and buy-now vs wait-for-the-M5-Ultra-Studio.
-
CUDA vs Vulkan for llama.cpp
CUDA still wins on NVIDIA, but in 2026 Vulkan matches it on token generation and runs on every GPU. Backends by vendor — CUDA, Vulkan, ROCm, SYCL — with real benchmarks.
-
How to Run AI Locally: The Complete Beginner's Guide
The on-ramp: run your first model in 10 minutes, then the whole landscape — hardware tiers from $0 up, model picks, quantization, backends, and fixes for the first errors you'll hit.
-
RTX Spark vs DGX Spark vs RTX 5090
RTX Spark is not a cheaper DGX Spark. Same family of silicon, different OS, memory speed and networking — we map all three options for local LLMs with real numbers.
-
Best build for local AI
A decision tree of best / runner-up / honorable-mention picks per budget, from a used RTX 3090 up to datacenter-class rigs.