Gains:
- Concrete understanding of the concept of model 'weight' and the implicit/explicit distinction
- See the difference between openweight and open source and their licensing pitfalls
- Ability to choose a closed, open or mixed approach with justification depending on the situation of the business
While getting to know the providers in the previous unit, there was a column called "weight type" in the table: closed, open, mixed. This single column actually hides one of the most strategic decisions of your model selection. Whether you will work with a closed or open weight model; It directly determines where your data will be processed, how your monthly costs will be shaped, your technical team needs and your legal responsibilities. By the time this unit is finished, you will have a concrete understanding of the concept of "weight", see the subtle but critical difference between open weight and open source, and be able to justify the right approach for your own situation.
What Exactly Is "Weight"?
An AI model is trained with a huge chunk of text while slowly adjusting the billions of numerical values (parameters) within it. All of these numbers are called model weights. Weights are where the model stores “everything it has learned”; You can compare it to the power of synapses in a person's brain. The model's grammar, world knowledge, and reasoning patterns are all coded in these numbers.
“Running” a model means loading these weights into a computer and running the text you enter over them. The critical question is: Who has access to these weight files?
- Closed-weight: Weights reside only on the provider's servers. You cannot access them directly; You use it over the internet via an interface (API) in a "send text, get response" manner. Anthropic Claude, OpenAI GPT, Google Gemini are like this. (API is an interface that allows a software to send a request to another software through a standard port and receive a response.)
- Open-weight: Weight files can be downloaded. If you want, you can install and run it on your own server. Some models of Meta Llama, DeepSeek, Mistral are like this.
Tip: Think of the closed weight as "rented apartment" and the open weight as "bought house". It is easy to move in a rented flat, the maintenance is not your responsibility, but the landlord sets the rules. Everything in the purchased home is yours, but renovations, security and bills are also your responsibility.
Open Weight is Not the Same as Open Source
There is a very common conceptual error here. "Open-weight" and "open-source" are often used interchangeably, but they are not the same thing.
- Open weighting means sharing only the learned parameters of the model. It is usually not explained how the model is trained and what data it is fed with.
- Fully open source ideally means that both the weights, training data, and training code are shared and can be used freely.
There is also the issue of licensing. Even if a model's weights are downloadable, its license may limit commercial use, impose additional requirements beyond a certain number of users, or prohibit competing product development. So "I can download it" and "I can use it as I wish" are different things.
Caution: Before putting an open weight model into your business, be sure to read its license. Don't be fooled by the word "open"; Some licenses require separate agreements above a certain number of monthly active users, and some restrict the use of the output for certain purposes. This is something you should ask your legal team about.
Comparison of Two Approaches
The table below puts the two approaches side by side on key dimensions.
Size
Closed Weight
Open Weight
Installation
None, ready interface
Server and technical team required
Where is data processed
On the provider's server
In an environment under your control
Initial cost
Low (pay as you go)
High (hardware, installation)
Cost at scale
increases with volume
Predictable with fixed hardware
Update/maintenance
at the provider
you too
privacy control
limited
full
Access to the most powerful model
usually here
Variable
In this table, there is no single answer to "which is better"; Each line is plus or minus depending on your priority. If you want the initial cost to be low, closed weight; If data must not leave the company even a millimeter, open weight comes to the fore.
Three Realistic Cases
Case 1 — E-commerce startup (closed gain weight). A five-person e-commerce team wants to produce product descriptions. They don't have technical teams, they don't want to manage servers, monthly volume is variable. They start with a closed-weight model, paying for what they use. In the first month, they process 1.2 million tokens and pay out around $15; As their business grows, the bill also grows proportionally. There is no server, maintenance or update problem. The right choice for this profile is clear.
Case 2 — Hospital (deficit gains weight). A hospital wants to summarize and code patient notes. The data is highly sensitive health data; should not leave the institution under any circumstances. They are building an open weight model in their data centers. They make a hefty initial investment in setup and hardware (say, a $40,000 server and two weeks of setup labor), but then they don't pay additional token fees per patient and the data never goes to an external server. The high volume and absolute confidentiality justify the initial cost.
Case 3 — Medium-sized accounting firm (hybrid approach wins). An accounting firm uses a closed-weight model from a ready-made interface for daily general correspondence and draft texts; because it is not worth the installation effort for these jobs. However, they outsource sensitive analysis involving customer financial data to an open weight model running on their own servers. Thus, they benefit from both convenience and privacy at the same time. We will return to this “mixed portfolio” approach in unit 9.
Weak Prompt / Strong Prompt: When Asking for Decision Support
Weak prompt:
Should I use open weight or closed weight?
Powerful prompt:
Your role: enterprise AI architect.My situation: We are a 12-person accounting firm, we have no technical team but we do have an external IT consultant. ~500,000 tokens per month business. Some of the data is customer financial data (sensitive), some is general correspondence. We have to comply with KVKK. Task: Compare closed, open and mixed approach accordingly. For each: suitability (1-5), estimated first year cost range, privacy note, main risk. Finally, give one suggestion and summarize the reason in 3 points.
The second prompt allows the model to produce a recommendation that really suits your business, as it gives all the boundaries required for the decision (team, volume, data type, legal framework).
Copiable Templates
1. Weight type decision:
My job: [DESCRIPTION]. Technical team: [YES/NO]. Monthly token volume: [NUM].Data sensitivity: [LOW/MEDIUM/HIGH].Budget structure preference: [LOW START / PREDICTABLE FIXED].Compare closed, open and mixed option according to this profile and give a single recommendation.
2. License check:
Simplify the license for this open weight model: [LICENSE NAME/TEXT].Answer these questions: Is commercial use allowed? Are there user/revenue limits? Can I resell the output? What are the uses it prohibits? If unclear, write "ask a lawyer".
3. Cost breakdown:
Break down the total cost of ownership for a bare-weight model: hardware, installation labor, electricity/hosting, maintenance, updates. Make a rough 3-year comparison with the closed weight alternative. My monthly volume: [NUMBER].
4. Privacy risk screening:
Trace where data goes from where to in the following workflow: [WORKFLOW DESCRIPTION].If I use closed weight, at what point does data go out? How does this risk change at open weight? Suggest a mitigation measure for each risk.
Common mistakes
- Mistaking the word “open” for “free and unlimited”: Open weight does not guarantee free or unlimited commercial use; license is decisive.
- Thinking open weight is "easy": Running an open weight model requires servers, technical staff and maintenance; Underestimating the installation is a common mistake.
- Not considering the closed weight data at all: In the closed model, the data goes to the provider; It's risky to start without sorting out in advance which data can go.
- Just looking at the initial cost: While open weight may seem expensive initially, it can be cheaper in the long run at very high volume. Look at the total cost over three years.
- The 'strongest pattern is always closed' motto: While this is true most of the time, it is not absolute; open weight models may be more than adequate for certain tasks.
In summary
- Model weights are billions of parameters that store the information the model has learned; The "off/on" distinction is about accessing these parameters.
- Closed weight: ready interface, low start, on data provider. The obvious weight: your own server, full control, high startup and maintenance overhead.
- Open weight is not the same as open source; License terms may limit commercial use.
- The right choice; Varies depending on team, volume, data sensitivity and budget structure. A hybrid approach makes sense for most midsize businesses.
Application task
Take the needs profile you created in the previous unit. Give it to an AI model using template 1 (weight type decision) in this unit. Evaluate the resulting recommendation yourself by matching it one by one with the seven dimensions in the table (installation, data location, startup cost, cost of scale, maintenance, privacy, access to the powerful model). If your evaluation differs from the model's recommendation, note in one sentence where you differ.
checklist
- [ ] I can explain the concept of "weight" in my own words.
- [ ] I can compare the advantages/disadvantages of closed and open weight in seven dimensions.
- [ ] I know the difference between openweight and open source and the licensing risk.
- [ ] I can justify the choice of closed/open/hybrid for my own business.
- [ ] I can track where data is going in a workflow and show the point of risk.