What exactly is an open-weight AI model? Not the policy argument that filled the tech press last week, but the practical question underneath it: is there a version of AI you can download, run on your own computer, and stop paying a monthly bill for?
The short answer is yes. The longer answer is where the money hides.
What is an open-weight AI model?
An open-weight AI model is one whose trained parameters, called the weights, are published for anyone to download. You can run it on your own hardware, inspect it, and adapt it. It is not a chatbot you log into. It is the engine, shipped without the car around it.
The weights are simply a very large file of numbers. Those numbers are everything the model learned during training. Hand someone the weights and you have handed them the finished brain.
Think of a restaurant. A closed model is the dish: you order, the kitchen stays hidden, you pay per plate. An open-weight model is the recipe and the ingredient list, handed to you at the door. You now own the method. You still need a kitchen, a cook, and gas.
One term to keep straight. Open weight is not the same as open source. Most releases publish the weights only. The training data and the full training code usually stay private.
How is an open-weight model different from the free version of a chatbot?
A free chatbot is a hosted service with a usage cap. An open-weight model is a file you host. With the free chatbot, the vendor owns the servers, the logs and the shut-off switch. With open weights, you own all three, along with the electricity bill and the person who keeps it running.
Three differences that matter to an owner:
--- Who holds your data. With a hosted service, your text travels to the vendor. With a model on your own box, it does not leave the room.
--- Who can change it. A hosted model can be updated or retired on the vendor's schedule. A downloaded model behaves the same way in three years as it does today, because it is a file sitting on your drive.
--- Where the cost sits. A hosted service bills per use. A self-hosted model bills per month whether you use it or not, because the hardware exists either way.
If the difference between running AI on your own machine and running it in someone else's data centre is the part you want to explore further, that decision is covered separately in our guide to local AI.
Why are open-weight models back in the news in 2026?
On 27 July 2026, Anthropic published a position paper in which chief executive Dario Amodei stated that the company "has never advocated for a ban on open-weights models", and that "open-weights models that don't have dangerous capabilities are a public good". The statement made news because many people had assumed the opposite.
Amodei set out two concerns. The first is that authoritarian governments may build models that surpass those of the United States for military or surveillance use, in which case whether the weights are public is beside the point, because the most dangerous models would be trained in secret. The second is that bad actors may misuse powerful models for cyberattacks or biological attacks.
Instead of bans, he proposed three measures: cracking down on industrial-scale distillation operations, mandatory safety testing for both open and closed models, and tighter chip controls.
Why a business owner should care about a policy paper. Because the supply of good downloadable models is a political question as well as a technical one. If you build a daily workflow on a model you downloaded, then licence terms and export policy are now line items in your risk register, not somebody else's problem.
The practical picture in 2026 is that the downloadable tier is genuinely strong. MindStudio's review of DeepSeek V4 places its agentic benchmark scores alongside GPT-5.5 and Claude Opus 4.7. Google's Gemma 4 shipped under the Apache 2.0 licence. Much of Mistral's lineup and many of Alibaba's Qwen models are Apache 2.0 as well.
How much does it cost to run an open-weight model yourself?
Cheap per token, expensive per month at small-business volume. One published analysis puts the break-even against a budget open-weight API at roughly 5.7 billion tokens a month on a single H100 GPU. Against a premium GPT-5-class API at around US$5 per million tokens, a separate estimate puts the crossover nearer 256 million tokens a month, and only at 60 to 70 per cent sustained GPU use.
Those numbers mean nothing until you convert them into your own work, so here is the conversion.
A million tokens is roughly 750,000 English words. A Hong Kong retailer handling 2,000 customer conversations a month, at about 1,500 tokens per conversation, consumes roughly 3 million tokens. That is a little over one per cent of the lower crossover point, and around one twentieth of one per cent of the higher one.
In other words, a typical small business will never reach the volume at which owning the hardware becomes cheaper than renting the service. If you want the arithmetic behind token counts, our explainer on AI tokens works through it.
There is one important exception, and it is about model size. NVIDIA Research estimates that serving a 7-billion-parameter model is roughly 10 to 30 times cheaper in latency, energy and compute than a 70-to-175-billion-parameter model on the repetitive, narrow tasks that make up most agent workloads. Smaller downloadable models also answer in 50 to 200 milliseconds when run locally, against hundreds of milliseconds for a cloud model. Gartner expects enterprise use of small, task-specific models to be three times that of large models by 2027.
So the honest framing is not "open weights are cheaper". It is "a small, narrow, downloadable model can be dramatically cheaper on one specific high-volume task".
Can a Hong Kong small business use these models commercially?
Usually yes, but the licence matters more than the benchmark. Apache 2.0 and MIT licences permit commercial use with almost no conditions. Meta's Llama 4 Community Licence permits commercial use but adds an acceptable-use policy, a naming requirement, and a threshold above which you must negotiate separately.
The three licence families you will meet:
--- Apache 2.0. Broad freedom to use, modify and distribute for almost any purpose. Google's Gemma 4, much of Mistral's lineup, and many Qwen models sit here. This is the most legally predictable option for a commercial deployment.
--- MIT. Similarly permissive. DeepSeek-R1 and Microsoft's Phi family are MIT licensed.
--- Community licences. Meta's Llama 4 Community Licence allows commercial use, but any derivative model must begin its name with "Llama" and display "Built with Llama" in product documentation, and commercial use above 700 million monthly active users requires a separately negotiated licence with Meta.
One data-residency point specific to Hong Kong operators. Running a model's weights on your own machine is a different question from calling that model's hosted API. The hosted DeepSeek API is operated from China, which raises data-residency questions for regulated or cross-border work. Downloading and self-hosting the same open weights sidesteps that particular issue entirely.
Whichever route you pick, under Hong Kong's Personal Data (Privacy) Ordinance you remain the data user responsible for personal data you collect and process. The model's licence does not transfer that responsibility. This is general information rather than legal advice, and a licensed adviser should review anything material.
What do people get wrong about free AI?
Four misconceptions come up in almost every conversation on this topic. Each one has cost somebody real money.
Misconception 1: open weight means free. The licence is free. The GPU, the electricity, the monitoring and the person who fixes it at 9pm are not. Budget the stack, not the download.
Misconception 2: open weight means private. It means private only if you host it yourself. Calling an open-weight model through somebody else's hosted API involves exactly the same trust question as any other vendor, because your text still leaves your building.
Misconception 3: open weight and open source are the same thing. Most releases publish the finished weights while keeping the training data and full training pipeline private. You can run and adapt the model. You cannot audit how it was built.
Misconception 4: if it is downloadable, it must be second rate. That stopped being true. Independent reviewers now place the top downloadable models alongside recent frontier releases on agentic benchmarks. The real remaining gap is support, integration and accountability, not raw capability.
When does an open-weight model make sense for a small business?
Three situations, and none of them is "to save money on a chatbot". Open weights earn their keep when data must not leave your premises, when one narrow task runs at very high volume, or when you need a model that will not be changed or retired underneath you.
A worked example. A Kwun Tong logistics firm with 11 staff processes about 4,000 delivery notes a month. Each one arrives as a phone photo containing a customer name, an address and a mobile number. The firm wants extraction only, no conversation. A small downloadable model on one on-premises box fits well, because the task is narrow, the volume is high and the personal data never leaves the office.
The same firm's two salespeople draft bilingual quotations a dozen times a day. That work goes to a hosted assistant instead, because the volume is trivial, the quality bar is high, and nobody wants to maintain a server so that a quotation reads well.
One company, two correct answers. That is normally how it lands.
Five questions before you download anything:
--- Name the single task. If you cannot describe it in one sentence, you are not ready to pick a model.
--- Count your monthly tokens. Multiply your monthly volume by roughly 1,500 tokens per interaction, then compare against the crossover figures above.
--- Price the whole stack. Hardware or cloud instance, storage, monitoring, and the hours of the person maintaining it.
--- Read the licence before the benchmark. A model that scores two points higher and carries naming obligations is the worse business choice.
--- Name the owner. Write down who is responsible when it stops working on a Tuesday morning. If that name is blank, choose a hosted service.
The takeaway
Open-weight models are a real option in 2026, and they are far better than most owners assume. What they are not is a way to stop paying for AI. The bill moves rather than disappears: away from a monthly invoice you can cancel, and towards hardware and staff time you cannot.
For most Hong Kong small businesses the sensible position is a mixed one. Rent hosted intelligence for varied, low-volume, judgement-heavy work. Consider a small downloadable model for the one repetitive, high-volume, privacy-sensitive job that never changes. Decide it task by task, not by ideology.
And if the whole subject still feels like it belongs to somebody with a server room, that is a reasonable reaction rather than a failing. We understand AI. UD stands with you.
Not sure which route fits your business?
Most owners do not need a verdict on open weights. They need to know which of their own tasks is ready for AI at all, and which is not worth touching yet. That is exactly what a readiness check answers, and we will walk you through it step by step, from the first assessment to a working deployment.
Reviewed by the UD AI team. UD has supported Hong Kong businesses with enterprise technology for 28 years.
