When discussing large language models, it’s easy to define “closed” or proprietary models. Models from the GPT, Claude, or Gemini families are clearly closed, as we don’t have access to their weights, training data, can’t deploy them on our own infrastructure, we aren’t free to fine-tune them, and how we can use them is strictly limited to the interfaces that OpenAI, Anthropic, or Google choose to publish.
Models become more open as the various limitations of closed models are removed. There isn’t a single definition of an “open model”; instead, there are degrees, depending on what they allow or not. The criteria are:
Weights: These are the values that make up the trained model. Having access to the model’s weights means having the technical ability to deploy the model for inference (running it) or modify it (fine-tuning, quantization, etc.).
Training data: Access to training data represents a big step in terms of transparency, allowing identification of, for example, copyrighted material or data biases, and making the model’s training process reproducible.
License: The license allows the author to specify who can use the model and how. Even if the weights have been published, certain licenses may restrict a model’s use, for example, by other companies, requiring royalty payments.
Let’s look at a couple of examples:
- Llama 3.1: The weights are open, while the training data isn’t public, and its license limits certain commercial uses and using the model’s responses to train other LLMs, among other things. For these reasons, some think that Llama models can’t be considered truly open.
- LM360: They make both the weights and training data of their models public and provide them under the Apache 2.0 license, a classic open-source license.
You can find a very comprehensive list of open models with their details here.
Pros and cons of open models
If we open Llama 3.1 70B and ask about the advantages of open models, it will generate a list including things like:
- Transparency: Having access to the model (and perhaps the training data) allows us to see how it works.
- Customization: The ability to make changes to the model, such as additional training, quantization, etc., to better adapt it to our use case.
- Community: We have access to improvements and modifications to the model created by the community, as well as shared knowledge.
- Cost: We have complete flexibility to host the model however and wherever we prefer, which can reduce costs.
- Privacy: The ability to run the model locally means we can ensure maximum privacy.
- Vendor independence: We won’t be dependent (if the license isn’t too restrictive) on a provider who might decide to discontinue the model we use, increase prices, etc.
We could do the same for disadvantages, but let’s focus on the two main ones:
- Lower capabilities: Since the most cutting-edge models are usually developed in a closed manner. The main exception in this case are the Llama models, which in their current 3.1 version are at a similar level to the most advanced models. But in general, it seems difficult for most companies to maintain a business model that, after investing the millions it costs to train an advanced model, then releases it for anyone to use for free.
- Complexity: If we decide to use an open model, more technical knowledge will be needed if we also want to manage the hosting ourselves; although this doesn’t have to be the case, as we can look for providers to manage this for us (for example, Groq).
All of these are very valid arguments.
In my personal opinion, however, the main reason I consider the use of open models to be the only valid option in many cases is control. Let me explain:
These days it looks like every new startup, service or product needs to be AI-related. In a way, this makes sense, since AI (LLMs in particular) is without a doubt the most important and transformative technology to appear in recent years. However, in most cases it feels like these products come from a decision to offer something AI-related first, while solving an actual real problem or being useful comes as an afterthought. The typical situation of a solution looking for a problem. Of course, in this environment, using the newest and flashiest LLM is of prime importance.
Of course, most of these companies or products will be short-lived. However, those that survive will necessarily not be based solely on showcasing a technological gimmick; they will have to deliver a legitimate value-adding product, which incidentally will make use of an LLM to deliver that value.
Successful products will use AI as one of the ingredients in their recipe for success. But for the product to be successful, all of its components need to be fine-tuned to work in unison, and that includes a specific LLM version. Once the product has been crafted, tested and delivered to users, it’s not trivial to switch to a new LLM (even if it’s a more advanced version of the same one) and expect the product to continue working in the same successful way.
However, if the product is based on a closed model, it will always be at the whim of the model provider who can discontinue or modify it. Even using third parties to host the models, like Amazon Bedrock, does not guarantee that the LLM we are using will still be available in 6 months.
This is why I wouldn’t consider building a real project on top of any model that isn’t open. The latest frontier models are exciting and great for experimentation and personal use, but I see the lack of control of using a closed model as too much of a risk for any serious endeavor.