What open-weight models actually buy you

You can download a capable model, run it on a machine you own, and never send a client’s file to anyone. That option is newer than most people realise, and it arrives with a permanent weather report attached. If you have read a headline suggesting that some lab or some government is about to take it away, the noise makes the decision feel riskier than it is.

The decision itself is duller than the politics. You are choosing between renting a model through somebody’s API and owning a file that runs on hardware you pay for, and the two differ on cost, quality, control, and who is able to interrupt you. This guide is for a solo operator or a small team working out where each one belongs. It is not for anyone training a model and publishing it: distributing a derivative is what switches on the redistribution conditions in these licences [3][4], and that job needs the licence texts and a lawyer rather than a guide.

Open weights means you get the file, not a set of freedoms

“Open weights” describes one thing only: the trained parameters are downloadable. What you may then do with them is set by a licence, and the licences differ enough that treating them as one category will eventually cost you.

At the permissive end, OpenAI released gpt-oss-120b and gpt-oss-20b on 5 August 2025 under “the flexible Apache 2.0 license”, with the weights “freely available for download on Hugging Face” [2]. That is an off-the-shelf licence rather than one the lab wrote for the occasion, and that is the entire difference between this paragraph and the two below it.

In the middle sits Meta’s Llama 4 Community License. It is free for almost everyone and conditional in ways a standard licence is not. If your product had more than 700 million monthly active users in the month before the release date, you “must request a license from Meta” [3]. If you distribute the materials you must “prominently display ‘Built with Llama’ on a related website, user interface, blogpost, about page, or product documentation”, and any model you train from them must “include ‘Llama’ at the beginning of any such AI model name” [3]. The user threshold will never touch you. The attribution and naming rules will, the moment you ship something built on it.

Google’s Gemma terms go further. Redistributors must include the use restrictions “as an enforceable provision in any agreement governing the use and/or distribution of Gemma or Model Derivatives”, must “cause any modified files to carry prominent notices stating that you modified the files”, and must ship a “Notice” text file pointing at the terms [4]. Google also “reserves the right to restrict (remotely or otherwise) usage of any of the Gemma Services that Google reasonably believes are in violation of this Agreement” [4]. That is a leash on a downloaded file, and it is the sort of clause people are surprised by later rather than earlier. Read the licence before you read the benchmark scores. It takes ten minutes and it is the only part of this decision that is genuinely irreversible once you have shipped.

The policy ask is narrower than the headline

The specific fear behind most of this, that open models are about to be legislated away, has a public answer from the vendor people usually assume is pushing for it. Anthropic published its position on 27 July 2026, and the first line of it is “Anthropic has never advocated for a ban on open-weights models” [1]. The same statement calls open-weights models without dangerous capabilities “a public good”, on the grounds that “they don’t cost anything besides the compute needed to run them, and they provide value to businesses, developers, and researchers”, and says plainly that “Protectionist bans would not address my most serious national security concerns” [1].

What it does ask for is three things: that the United States not sell advanced chips or chipmaking equipment to China, that “we should crack down on industrial-scale distillation operations”, and that “all sufficiently capable models, open and closed, should go through mandatory safety testing” [1].

Notice what none of those do to you. Chip export rules act on who can train large models and how cheaply, which reaches you eventually through the price and pace of new open releases, not through your right to run the one already on your disk. Mandatory pre-release testing acts on the release, not on the deployment. The distillation ask is the only one aimed at a practice rather than at hardware, and, as the next section covers, the version of it that already applies to you is not in any statute.

The durable habit here is smaller than the topic. When a policy story touches your stack, find the primary document and read the actual ask before you change anything. The gap between “lab wants open models banned” and the three asks above is the gap between rebuilding your stack and doing nothing.

The distillation clause that binds you is one you already agreed to

If you use a hosted model, you accepted terms. Anthropic’s Commercial Terms of Service, effective 17 June 2025, say a customer may not “access the Services to build a competing product or service, including to train competing AI models or resell the Services except as expressly approved by Anthropic”, nor “reverse engineer or duplicate the Services” [8].

Read that against a plan people reach for often: use a strong hosted model to generate a few thousand example answers, then fine-tune a small open model on them so the cheap model does the job. That is training a model on another vendor’s outputs, and whether a given version of it falls inside a clause like the one above is a question for the terms you signed and for someone qualified to read them, not for a confident answer in a guide. The point is that this is where the distillation question actually lands on a small operation. Not in an export-control fight, but in a contract you clicked through, on a plan that sounds like a clever cost saving.

Using a hosted model to write your code, your drafts, your evaluation cases, and your documentation is not that. Building a competing model is. Most work sits nowhere near the line. If yours does, get it checked before you spend a month on it.

The interruption to plan for is a retirement notice

The thing most likely to break your setup this year is not policy. It is a model you rely on being switched off on a schedule the vendor sets.

Anthropic publishes the lifecycle in four states: active, legacy, deprecated, and retired, where a retired model “is no longer available for use” and “Requests to retired models will fail” [7]. Customers with active deployments get “at least 60 days’ notice before model retirement for publicly released models”, and the guidance is to “migrate all usage to a suitable replacement before the retirement date” and to test your applications with the new models “well before the retirement date” [7]. That is a fair policy, clearly documented, and it is still a clock you do not control.

A file on your disk has no such clock. This is the honest core of the open-weights case, and it is narrower than the marketing around it. You are not buying freedom, because the licence still governs you. You are not usually buying savings, as the next section shows. You are buying continuity: a model that behaves the same way in two years, for work where behaviour drifting under you is the expensive failure.

The rest is preparation you should do whichever way you go. Keep your prompts in plain files rather than scattered across tool configurations. Keep twenty real jobs with answers you already accepted, so that swapping a model is a test rather than a feeling. Both habits turn a forced migration into a Tuesday afternoon instead of a bad week.

Price the job before you price the hardware

Cost is where the open-weights argument is weakest for small operations, because hosted small models have become very cheap. On current list prices, Claude Haiku 4.5 is $1 per million input tokens and $5 per million output [5], and OpenAI’s GPT-5.6 Luna is $0.20 per million input and $1.20 per million output [6]. The flagships are a different world: Claude Fable 5.1 is $10 and $50 per million [5], and GPT-6 Astra is $10 and $50 [6]. So Astra costs 50 times as much per input token as Luna [6], and Fable 5.1 costs 10 times as much as Haiku 4.5 [5]. The routing decision inside one vendor’s price list is usually worth more money than the open-versus-hosted decision between vendors.

Work out the actual number for one job before arguing about architecture. Run the volume you really have, at the price of the cheapest model that passes your twenty test cases.

Against that, price the other side properly. Running gpt-oss-20b locally needs 16GB of memory and running gpt-oss-120b needs 80GB [2]. The first is a well-specified laptop. The second is not, and the gap in what you get for the difference is real: OpenAI puts gpt-oss-120b at matching or exceeding o4-mini on general problem solving and tool calling, while gpt-oss-20b “matches or exceeds OpenAI o3-mini on these same evals” [2]. Then add the part nobody costs: somebody keeps the machine patched, the runtime current, and the thing answering at 7am. On a bill of a few dollars a month, that somebody is you, working for free.

Running your own weights is a control decision

There are good reasons to hold the file, and they are all about control rather than money.

Data that cannot leave your premises is the clearest one. If a client contract or a regulator says material stays on your hardware, a local model is not a preference, it is the only option that complies, and no price comparison applies. Work that must run without a network is the second. Behaviour that must not change is the third: a classifier that has been tuned against your own examples and needs to give the same answer next year has no equivalent on a hosted endpoint with a retirement date [7].

Volume is the fourth, and it arrives later than people expect. Plug a genuinely large number into the calculator above and the hosted bill eventually crosses the cost of hardware and attention. Until it does, renting is cheaper than owning and much cheaper than administering.

A sane shape for most small operations is a hosted model for the varied thinking work, a cheap hosted model for the high-volume mechanical work, and a local model only where one of those four reasons applies. That is a stack, not a side. The mistake is picking a team.

What still goes wrong

The quality gap is real and it is easy to talk yourself out of. The model that fits in 16GB is benchmarked by its own maker against o3-mini [2], not against a flagship at $10 per million input tokens [5][6], and the tasks where that difference shows up are the ones you care about: long documents, tangled instructions, anything where being confidently wrong costs a client relationship. Test on your own work rather than on published scores, and be willing to conclude that the cheap local option is not good enough yet for the specific job you had in mind.

Licences are attached to versions. Meta’s user threshold is measured “on the Llama 4 version release date” [3], which tells you the document was written for that release rather than for whatever follows it. If a clause is load-bearing for your business, say Google’s right to “restrict (remotely or otherwise) usage” of Gemma Services [4] or Meta’s naming and attribution rules [3], re-read it at every upgrade instead of assuming it carried over.

And the policy picture stays unfinished. Anthropic’s stated asks cover chip exports, industrial-scale distillation, and mandatory testing of sufficiently capable models [1], but a stated position from one company is not law, and none of it is settled. The reasonable posture is the one this guide has argued for throughout: keep your prompts portable, keep your test cases, know which licence you are operating under, and treat any headline about open models being banned as a prompt to read a primary document rather than to rebuild anything.