Vetted Consumer is an independent publication about the hardware that runs open-weight language models locally: unified-memory machines such as the Mac Studio and the Strix Halo mini PCs, graphics cards judged through the local-LLM lens, AI mini PCs and servers, edge accelerators, and the software that runs them. It launched in May 2026 and is written and edited by Thomas Newkirk, an IT manager in Dayton, Ohio. It is non-commercial: no affiliate links, no sponsorships, no review units, and no payment from any company we cover.
What we're for
There's no shortage of tech reviews, there's a shortage of clear decisions. For any given product, you can find a dozen reviews, a hundred Reddit threads, and a pile of conflicting benchmark numbers. Vetted Consumer exists to do that reading for you and answer the only questions that matter when you're about to spend money: Is this the right product for me? Who should buy something else? And what's the catch nobody mentions until after you've bought it?
How we work We're a synthesis publication, and we're upfront about exactly what that means, read How We Research for the full picture. In short: we cross-reference multiple independent reviews and verified owner reports, we name and link every source, and we clearly label anything we've observed first-hand. We don't run our own lab tests, and we never pretend otherwise.
Independence Vetted Consumer is reader-funded and supported by affiliate links, which never influence our verdicts. We say "skip it" as often as "buy it." See our affiliate disclosure.
Who's behind it Thomas Newkirk, an AI enthusiast who is loving every second of this marvel technology. Reach me at [email protected]. I correct mistakes; if something here is wrong, tell me.
Our own hardware
Most numbers on this site are other people's measurements, attributed and linked: owner reports, independent labs, and primary documentation. The site does own one test machine, a Mac mini with an Apple M4 and 32GB of unified memory, the entry tier we write about most. Where we publish a figure measured on it, the article says so, names the runtime and version, and gives the date, so a first-party number is never confused with a ceiling or with someone else's result. Everything else stays attributed to whoever measured it. The full method is on the methodology page, and mistakes go on the corrections log.
First-party numbers, measured 12 September 2026
Mac mini, Apple M4 (10-core GPU), 32GB unified memory, macOS 26.6.2. Runtime: llama.cpp build b10929, the official macOS arm64 release, Metal, every layer on the GPU. Method: llama-bench, prompt processing at 512 tokens and generation of 128 tokens with 3 repetitions each; prompt processing at 8,192 tokens with 1 repetition; generation of 128 tokens with the KV cache pre-filled to a depth of 8,192 tokens (the -d flag) with 2 repetitions. Model files are the Q4_K_M GGUFs published by unsloth on Hugging Face, checked against the publisher's SHA-256 before running. Nothing else ran on the machine.
| Model (Q4_K_M) | File | Prompt tok/s (512) | Generation tok/s | Prompt tok/s (8,192) | Generation tok/s at 8K depth |
|---|---|---|---|---|---|
| Qwen3 4B | 2.49GB | 416.6 | 36.2 | 299.9 | 24.6 |
| Qwen3 8B | 5.02GB | 223.8 | 19.7 | 174.5 | 16.1 |
Standard deviations across repetitions were under 0.4 tok/s in every cell. For comparison, the site's calculator gives bandwidth ceilings of 25.0 tok/s (8B) and 50.0 tok/s (4B) with an empty context and 20.4 and 35.9 at 8K depth on this machine; the measured figures land at 70 to 80% of those ceilings in every cell, the normal gap between a bandwidth ceiling and a real runtime. These two rows now appear in Can I Run It, the Quant Picker and the MCP server as "Measured on our Mac mini", separate from owner reports. The raw llama-bench JSON is kept in the site's repository.