Open weight vs open source: the distinction your board is about to get wrong

This week a team at Meituan released LongCat-2.0, a 1.6-trillion-parameter model, and the word that traveled with the headline was “open source.” Within a day it was in every AI news roundup, MIT license attached, sitting on the leaderboard, and by Wednesday someone on an exec team somewhere had forwarded the link with a one-line note: can we just run this instead of paying the API?

It is a fair question. It is also the moment where a small vocabulary mistake turns into a real one. Because the model was announced as open source while, per MarkTechPost’s launch writeup, “local self-hosting is not yet possible, since weights remain pending.” The label arrived before the thing the label describes. That is not a scandal. It is just how loosely the word “open” gets used now, and it is exactly the confusion that lands a bad assumption in a board deck.

The 1.6-trillion-parameter model that shipped with “open source” on the label

Here is what actually landed. LongCat-2.0 is a mixture-of-experts model, and the specs are genuinely impressive.

Native one-million-token context, MIT license, and, notably, trained on a 50,000-card domestic ASIC cluster with no Nvidia hardware in the loop. By the July 9 news roundups the weights had reached Hugging Face, so it did become self-hostable. It was not the only “open” release either. This was a busy week, the kind where the release feed reads, in one tracker’s words, like “a quiet day after a week of nonstop model releases.”

So the board’s instinct is right that something real shipped. The gap is in what “open source” is quietly promising them.

Open-weight vs open source: what the label actually promises

The clean version fits in two sentences. Open weight means the trained parameters are published, so a team can download the model, run it, and fine-tune it. Open source, in the sense the Open Source Initiative uses, means the weights plus the training code plus enough data documentation to rebuild the model from scratch and audit how it was made.

Almost nothing marketed as open source today clears that second bar. As one plain-spoken explainer put it this month, “almost every model marketed as ‘open source’ today, Llama, DeepSeek, Qwen, Gemma, is actually only open-weight.” They ship the finished parameters, not the recipe. That is not a knock on them. Open weight is enormously useful. It is just a different thing than what “source” implies to anyone who has shipped software.

The part that matters for a self-hosted decision is that neither label tells you the license terms, and the license is the thing with teeth.

That bottom row is worth a second look. OpenMDW-1.1 is a license class built this year specifically for model weights, covering the parameters, the architecture docs, the training code, and the generated outputs under one framework. Poolside shipped its Laguna XS 2.1 coding model under it, retiring the old version on July 9. The ecosystem is inventing new legal vocabulary because “open source” and “open weight” were never precise enough to describe what these releases actually are.

Three questions the board asks before trusting open weight models

When leadership asks “can we just use this,” three real questions are hiding inside it. Calm, specific answers are available for all three.

Can we use it commercially without tripping a threshold later? For an Apache or MIT model like LongCat-2.0, yes, cleanly. For a custom community license, read the fine print on user caps and regions before building a business on it. The teams that get burned are the ones who assumed “open” meant “free of obligations” and discovered the cap mid-fundraise.

Does self-hosting it make us private and compliant by default? No, and this is the one worth slowing down on. Running an open-weight model in-house changes who holds the data and which jurisdiction can reach it. It does not, by itself, satisfy the EU AI Act. The Act’s open-source exemption for general-purpose models is narrow: the license has to permit use, modification, and redistribution, the model cannot be monetized in certain ways, and the parameters and architecture have to be public. Even when a provider qualifies, it still owes an EU copyright policy and a training-data summary, and the Commission’s enforcement powers, with fines up to 3% of worldwide turnover or 15 million euros, begin on August 2.

Does “open source” mean we can audit and reproduce it? Usually not. Open weight lets a team inspect and test behavior, which is real and valuable. It rarely lets anyone reconstruct the training data or reproduce the model, which is what a true audit would require. For a regulated shop, that distinction is the difference between “we tested it” and “we can prove how it was built.”

Open weight does not mean open source, does not mean free, and does not mean license-clean. It means the parameters are downloadable. Everything else is a separate question.

The one-minute answer on whether to self-host this week’s release

If the board gives one minute, here it is. A strong new open-weight model landed, and if it carries an Apache or MIT license, the commercial terms are genuinely clean. Self-hosting it can be the right call for data control and, above a real volume line, for cost. But “open source” on a headline is a description of vibes, not of rights or of compliance. The rights live in the license file. The compliance lives in the EU AI Act’s actual conditions. And the cost case still depends on utilization, not on the download being free, with H100 rental holding around $3.35 an hour this week and doing nothing for a team running its GPUs at 9%.

None of that argues against local. It argues for reading one license file before anyone promises the board a number.

The license shift worth tracking into August

The thing to watch is not the next 1.6-trillion-parameter release. It is August 2, when the EU AI Act’s enforcement powers switch on and the open-source carve-out stops being a talking point and starts being a test a provider either passes or fails. New license classes like OpenMDW-1.1 are appearing precisely because the old words could not carry that weight. The teams that stay calm through it will be the ones who already learned to ask “which license” instead of “is it open.” That question was always the real one. This week just made it easier to see.