What 'open-source AI' actually means — licences, weights, MoE
Openness comes in tiers, and the licence line — not the word 'open' — sets what you can do
The three lines
- 'Open-source model' is not one thing — openness runs from weights-only to full disclosure
- The practical test is the licence: Apache 2.0 means commercial freedom; research-only doesn't
- MoE — waking only some experts per query — is now the standard for large open models
Key questions
- Can I just use an open-source model commercially
- Depends entirely on the licence. Apache 2.0 or MIT: yes, including commercial services. Custom licences (like Meta's Llama terms): yes with conditions, such as user thresholds. Research-only licences: no. Reading one licence line beats reading ten headlines.
- What's the difference between open weights and open source
- Releasing the trained model file is open weights; strict open source also discloses training code, data and methods so third parties can reproduce it. Most 'open-source models' in the news are the former. If you only need to run the model, weights suffice; if you need to audit or reproduce it, the difference is everything.
- Why is everyone using MoE now
- It decouples knowledge size from compute cost. Total parameters grow to hundreds of billions while each query wakes only the relevant expert blocks — big-model knowledge at mid-model cost. DeepSeek's releases and Korea's K-EXAONE 2.0 (37B active of 750B) both ride this design.
"Open-source model released" ran twice in headlines this week alone — DeepSeek, then LG's K-EXAONE. But two releases wearing the same adjective can permit completely different things. This is a reference document for reading open-model news: the openness tiers, the licence test, and the architecture word that keeps appearing. Come back whenever the next release lands.
1. Tiers of open — ask what was actually released
The common misreading is "open source = everything public." In practice there are layers. Open weights means the trained model file is downloadable and runnable — most "open-source models" in the news are exactly this. Strict open source (as institutions like OSI argue) additionally discloses training code, data and methodology so third parties can reproduce the model — and very few frontier-scale releases clear that bar.
Which tier matters depends on who you are. A developer who just runs the model needs weights. A researcher or regulator auditing bias and capability needs reproducibility — which is why "who tests open models" surfaced as the gap in this week's White House framework debate.
2. The five-box test for open-model news
| Box | Question | Example |
|---|---|---|
| Licence | commercial use allowed? | Apache 2.0/MIT: yes · research-only: no |
| Scope | weights only, or code+data too? | mostly weights only |
| Architecture | dense or MoE? | recent large releases: mostly MoE |
| Active parameters | the real inference-cost signal | K-EXAONE 2.0: 37B of 750B |
| Benchmarks | self-run or independent? | at release: almost always self-run |
MoE (mixture of experts) in one line: the model is divided into many "expert" blocks, and each query wakes only the relevant few — so a 750B knowledge store computes like a ~37B model. That decoupling of size from cost is what made frontier-scale models runnable outside big-tech datacenters, and it is the technical engine behind the open camp's rise.
3. What is still open
The next battleground for open models is verification and liability: closed models can accept pre-release government testing (this week's 30-day framework), but once weights are public that concept has nothing to grip. How regulation handles the gap is the second half of 2026's question. This week's cases run in "LG open-sources a 750B model" and "Thirty days before launch"; general release-reading method in "What 'frontier model' means." This document updates as the landscape does.