Apache 2.0 Hands You the Weights and the Liability

A download with nothing in front of it

On 15 July 2026, Thinking Machines Lab published Inkling, a 975-billion-parameter open-weights multimodal model with 41 billion active parameters, and put the weights on Hugging Face behind nothing at all. No request form. No approval step. No waiting list. An engineer at a mid-size European firm can have the files pulling down before the coffee goes cold, and that is exactly what the lab intended.

What did not travel with the weights is the part that should hold an owner's attention. The company's own announcement blog names no licence. The grant that governs what you may do with those files lives somewhere else, and it is more permissive than most people assume when they hear the phrase "open weights". Read the two documents Thinking Machines did publish, side by side, and a clear shape emerges. The licence gives you everything and protects you from nothing.

Where the licence actually lives

Inkling is licensed under Apache 2.0, and that is established by the model card. The Hugging Face model card frontmatter reads license: apache-2.0, and the card text says the same thing in prose. That is authoritative. It is also the only place a reader will find it, because the company blog that introduced the model to the world is silent on the question.

Apache 2.0 carries no field-of-use restrictions. It does not tell you which industries you may deploy in, which topics you may generate about, or which customers you may serve. It is a permissive grant, written for software, and it does what permissive grants do. It hands over the rights and it disclaims the warranties. Nothing in it says a word about the content the weights produce once they are running on your hardware.

The gap between the blog and the model card deserves a sentence of its own. If someone in your business has already written a procurement note that says "we checked the vendor's announcement", that note checked the wrong document. The licence was never there to check.

The policy that reaches the weights

Thinking Machines separately publishes a Model Acceptable Use Policy, and it reaches the weights themselves. It is not modest in what it asks of you. It states plainly: "If you permit others to use the Model Materials through your products or services, you are responsible for their compliance as well." Read that as written. Wrap Inkling in a product and sell it, and the policy puts your customers' behaviour on your side of the ledger.

Now look at what the policy leaves out. It never uses the word "license". It never mentions Apache. It sets out no enforcement right, so it describes no mechanism by which the lab could act on a breach. It contains no termination clause, so it describes nothing that could be taken back from you. And it contains no indemnity, so nothing in it commits the lab to standing behind you when a use goes wrong.

We are not going to tell you whether that document binds you, and we are not going to tell you it is void. That is a question for your counsel and it will turn on facts we do not have. The structural facts need no lawyer, and they are the ones that matter for planning. The policy is not the licence, and there is nothing inside it that pays for a mistake.

A model trained to refuse less

This matters more for Inkling than for a typical open-weights release because the lab deliberately reduced the model's willingness to decline. Thinking Machines says so in its own words: "we trained Inkling to answer directly on topics that may be subject to censorship". That is a design choice, stated openly, and there is a serious argument behind it. It also means the safety behaviour many operators quietly assume is baked into any frontier model has been dialled down on purpose in this one.

The lab is candid about what survived the training. It concedes an "occasional tendency to comply with role-play and indirectly framed prompts concerning harmful topics". MarkTechPost covered the release on 15 July, and Artificial Analysis has treated Inkling as a leading US open-weights model. The attention is real, which means the number of businesses about to run these weights inside products aimed at European consumers is real too.

Our read: set the licence, the policy and the training choice next to each other and the liability picture is unusually clear. If you self-host Inkling and it produces something unlawful in the EU or the UK, the exposure that lands on you is statutory. It arrives from your own regulator, under the rules that already govern what your business publishes and distributes. It does not arrive from Thinking Machines, and the acceptable use policy is not a shield you can hold up. That policy is a statement of the lab's hopes for how its model gets used. Your regulator has not read it.

The moderation layer is not in the download

The instruction is small enough to run this week. Before you self-host any open-weights model, read the licence and the acceptable use policy as two separate documents, and ask one question of each. What does this protect me from?

For Apache 2.0 the answer is nothing, by design. It is a grant of rights rather than a warranty of outcomes, and it was never built to carry the question you are putting to it. For an acceptable use policy with no indemnity, the answer is also nothing. Two documents, one answer, and the answer comes back the same both times.

What follows from that is a line in a budget. The output of a model you host is your publication. If it says something defamatory about a customer, your business said it. If it produces material that breaks a rule your sector already lives under, the regulator's letter carries your name on the envelope. So budget for the moderation layer you did not think you were buying. Inkling is free to download and the filtering you will need around it is not, and the distance between those two numbers is the real price of the release.

Read next: Moonshot's Biggest Model Has No Thinking Dial  |  More Than 15 Partners, and No Published Rules