Who is local AI really for?
Matthew Berman showed 8 AI tools in this video on . 7 are verified available in the EU. Each row below jumps to the exact second. Part of week 2026-43.
Tools shown in this video, with timestamps
Sponsor segment: a development environment that lets you bring any agent (Claude, Codex, Gemini) and run them in parallel, reviewing with IDE tools like the debugger.
Matthew invested in the company; calls it the easiest product for running local AI, though still not dead simple due to hardware and model differences.
A compact open-weights model that fits on an RTX 4090/5090, running at roughly 25-70 tokens per second. On the cost chart it comes to about $144/month amortized with an intelligence index of about 34.
Coding-specialist open model running at 50-130 tokens/s on an RTX 6000 Ada and 95-245 tokens/s on an RTX Pro 6000 Blackwell, single-threaded.
Described as a roughly 2-year-old open-weights model that still runs at 75-190 tokens per second on an RTX Pro 6000 Blackwell.
Cutting-edge open-weights model that can't run on these consumer or prosumer devices unless heavily quantized, which costs some intelligence.
The 256GB/512GB configurations (about $10-15K) can run frontier open models at 20-55 tokens/s, but they currently have multi-month waits and are only usable at best.
Newly released Anthropic model that costs about $27/month and runs at about 235 tokens/s. It scores roughly 10 points higher than Qwen 3.8 27B on the Artificial Analysis index, and closed models dominate the Pareto frontier.
