Point of view
The local AI manifesto
AI is better at home. Not as a nostalgia project, and not as a rejection of the cloud - as a practical claim about who controls the model you depend on. This is what Tokenstead believes, and what we publish accordingly.
Own the machine.
A model you can run is a model nobody can switch off. Vendor access can be revoked, prices can move, and a policy change can put a frontier model behind a desk you cannot reach. Weights you have downloaded and hardware you have paid for are the version of AI that stays yours.
Open weights are a public good. Read the license.
Open is a claim; the license is the fact. MIT and Apache 2.0 are different from a research license, which is different from open weights with a carve-out for large companies. We check the license on every model page because the license decides whether you can build a product on it, not just run it at home.
Cost is part of the argument.
The cloud bill is a subscription to someone else's roadmap. Local hardware is a capital decision: a one-time cost against electricity, noise, and the models that actually fit. We publish budget builds with real parts and checked prices, and own-versus-rent math, because a good argument has to survive the invoice.
Verify the numbers. Keep the receipts.
Benchmarks are cheap to claim and expensive to reproduce. We favour measured runs, named hardware, stated quants, and linked sources over leaderboard screenshots. A number without a setup is marketing. The site's job is to show the setup.
Tools over hype.
A harness that makes a 30B model usable on a 12GB card matters more than another point on a frontier chart. New agent harnesses, local runtimes, quant tools, and evaluation utilities are how the community turns weights into work. We track the ones worth your attention and benchmark the harnesses people actually run.
Policy is infrastructure.
Export controls, AI acts, liability rules, procurement, and license enforcement decide what can be downloaded, run, and shipped. A self-hoster cannot opt out of that layer. We treat policy stories as operating news, not politics as spectacle, and explain what each change means for a rig on your desk.
Build in the open, with company.
Local AI is a practice, not a product. The useful knowledge is in the rigs, the power bills, the quantization choices, the failed runs, and the person one tier behind you. Share the numbers, ask the question, answer someone else's. That is why the community exists.
What that means for what we publish
We publish
- Opinion and analysis on where local AI is going, not press releases.
- Hardware cost math and budget builds with real parts and checked prices.
- Policy changes and what they do to open weights and self-hosting.
- New harnesses, runtimes, and tools that change what fits or what is usable.
- Community rigs and measurements with the settings attached.
We do not
- Run a raw model feed. Plenty of sites list every release; that is not the value.
- Take vendor placement or rank anything because someone paid.
- Call a cloud-only model local, or hide a license restriction.
- Publish a speed claim without saying what hardware and quant produced it.
This is a living document. If a principle is wrong or a claim does not hold up, tell us.