AI Guides › Trend Watch

The Difference Between An Open Model And An Open Company

By Nigel Guy · 2 min read

"Open" is one of the most stretched words in this industry. A company releases some model weights and gets described as open; another publishes a research paper and gets the same label; a third has "open" in its name and little else. People then carry that single word across to everything else the company does — its data practices, its governance, its future behaviour — as if one release settled all of it.

The Openness Ladder

Openness comes in layers, and a release can sit on one rung without touching the others. When you see the word, work out which of these it actually covers:

1. Open access — you can use the model through an interface or API, usually free or cheap. This is availability, not openness in any deeper sense.
2. Open weights — the trained model files are downloadable, so you can run it yourself. What you're allowed to do with them depends entirely on the licence.
3. Open licence — the terms genuinely permit use, modification, and redistribution, including commercially, without awkward carve-outs. Many "open" releases have restrictions here worth reading.
4. Open recipe — training code, methods, and meaningful detail about the data are published, so others could understand or reproduce the work.
5. Open organisation — the company is transparent about governance, decision-making, safety practices, and how it handles your data across all its products, not just one release.

A release can be high on rung two and low on rung five. That's not a contradiction; it's just a company making one specific thing available.

The questions to ask

• Which rung is this? Name it precisely rather than accepting the general word.
• What does the licence actually say? Look for limits on commercial use, user-count thresholds, acceptable-use clauses, and whether terms can change.
• Does this release tell you anything about the company's other products? Usually not. Their hosted service may have entirely different data terms from the downloadable model.
• What would you need openness for? Running locally for privacy, customising for a niche task, or simply avoiding lock-in each need a different rung.

What to skip

• Skip treating "open" as a virtue signal. It's a description of terms, and terms can be generous or narrow while using the same word.
• Skip the tribal debate about whether a given company "counts" as open. Your question is whether the specific release meets your specific need.
• Skip assuming closed means untrustworthy or open means safe. Each carries different risks; neither label is a verdict.

Guardrails

• Licences vary widely and change between versions — read the one attached to the exact release you're considering, not a summary of it.
• If you're relying on a licence for commercial work, get it checked by someone qualified; this guide is a reading aid, not legal advice.
• Running open weights yourself moves responsibility for security, updates, and misuse onto you. That's a real cost, not just a freedom.
• A company's past openness doesn't bind its future releases. Evaluate each one on its own terms.

All 751 AI guides · JulieMango plans from £17/mo