Desk
One person, one box, an easy runner. Fine for learning and private drafts. Not what you put behind a classroom.
Operators
Weights are not a system.
A downloadable model with no runtime, no evals, no logs, and no swap path is a souvenir. The civil-liberty claim holds only if a lab, school, or shop can run, measure, and replace the system on machines they control.
An operator is the person who keeps that system alive in use. They are not the lab that trained the weights. They are not the end user in the chat box. Closed APIs collapse the job into an account. Open weights create the job.
Stack
The model is one input
Work from the metal up. Defaults change. The contracts should not: hash the files, serve through a stable socket, judge the output, watch the traces, and keep a path to swap the weights.
Weights
Files on disk you can hash: parameters, tokenizer, template, adapters, quant recipe. A name in a catalog is not possession.
Engine
The runtime that turns those files into tokens on this machine. Desk tools optimize for minutes-to-first-token. Plant tools optimize for concurrent load.
Gateway
One stable endpoint in front of several engines so the application does not care which model is hot tonight. Failover dies when the names do not match.
Harness
Prompts, tools, control flow, memory. The scaffolding that makes the model swappable. Gains often come from editing this layer, not the trophy weights.
Eval and trace
A release you cannot fail is a demo. Traces, scores, and a red-team pass are how an operator knows Tuesday got worse.
Identity, secrets, retrieval, and a teaching surface sit around these five. Skip them and “local AI” is an open port or a hobby only its author can use.
Posture
Desk, shop, plant
Not one true stack. Three honest sizes.
Shop
Compose: engine, gateway, identity, traces, backups. This is the missing public default for a lab or school. Upgrades and drivers are real work.
Plant
A fleet, shared weights, routing, GitOps, private ingress. What product traffic looks like when it is not a slide. Unit economics and liberty are separate questions. Ask both.
Gaps
The engines are good enough
The missing public goods are in the middle: contracts, evals, runbooks, and people. Heiwig will not compete with inference engines. It will work the layer that makes those engines usable as liberty rather than as a weekend project.
- A named desk / shop / plant contract with pinned versions, one backup drill, and one eval that must pass.
- Runbooks for swapping weights without rewriting the application.
- Eval packs for real jobs — classroom, literature notebook, shop assistant — not only leaderboard tasks.
- Incident language for local systems: what you log, retain, and refuse to serve.
- People who can read a latency graph and a quant choice, not only a prompt.
Translation
Ethics as operations
Run on machines you control
Engine, weights, and a network you can name.
Study and fork
Logs, evals, and a swap path.
Safety is not a monopoly
Test this build. Do not outsource judgment to a vendor policy.
Release is a one-way door
Version the weights and the harness. You will not recall a USB stick.
If the gateway can only speak to one closed API, you are not operating open-source AI. You are theming a tenant portal. The written standard is on Ethics.
Heiwig
Own the missing middle.
A published desk / shop / plant matrix. Eval packs for the work we already named. Teaching that can bring a shop stack up, fail an eval on purpose, and swap a model. A working group that treats harness, gateway, and eval as public infrastructure.
Those outputs will be posted when they exist. This page is the contract, not a fake product list.