The Local AI Rig Sheet — 20 Free Local AI Models Sorted by the Machine You Already Own (8 / 16 / 24 / 32 / 64+ GB)
The companion to the full video: 20 free local AI models sorted into five machine tiers by RAM — 8, 16, 24, 32 and 64+ GB — plus the Mac, Windows and Linux hardware that runs each tier. Tiers here are set by the VERIFIED Ollama download size read from the live library page, never by the marketing parameter count in a model's name, because putting a model one tier too high tells you to buy hardware you never needed. INSIDE: (1) THE TIER MAP — installed RAM vs usable budget vs model size band, and why Apple Silicon only hands the GPU about 75% of unified RAM while macOS takes another 3–4 GB before you budget anything. (2) ALL TWENTY MODELS with exact download size, context window, the one thing each one is actually good for, and the exact `ollama pull` command — from GLM-OCR at 2.2 GB that turns scanned contracts into searchable text, through Qwen3.6 27B and Qwen3-Coder 30B on a 24 GB machine, up to Qwen3-VL 235B at 143 GB. Plus every pull command collected in one copy-paste block. (3) THE BUY TABLE — Mac, Windows and Linux options at every tier with published prices, from a used M1 Air around $500 and a used RTX 3090 around $700 up to an RTX PRO 6000 and a DGX Spark. (4) THE SPEED FORMULA — decode speed is bound by memory bandwidth, not compute: bandwidth in GB/s divided by model size in GB is your token ceiling, with published bandwidth for seven parts and a worked example. (5) THE FOUR-STEP SETUP, ending with the one line that points the AI app you already use at your own machine: http://localhost:11434/v1. (6) AN HONEST-LIMITS PAGE — the frontier models no consumer machine runs, why '1M context' is a spec sheet and not free memory, and why local is not free but prepaid. SOURCING: nothing in this sheet was benchmarked by Hyperautomation Labs. Every tokens-per-second and benchmark figure is a published third-party number printed next to the source that published it. Model sizes read from the live Ollama library on July 25, 2026; hardware prices as published the same day and subject to change. Two models on the sheet (LFM2.5 and GLM-OCR) carry no licence claim because their terms could not be verified before publication. Independent field sheet from Hyperautomation Labs — not affiliated with Ollama, Alibaba, Google, DeepSeek, OpenAI, Cohere, Apple or NVIDIA.
Free. No spam. Unsubscribe anytime.