This is the introductory post in a five-part series exploring open-source models. We'll cover what open-source models are, when you should use them, and how they compare to closed-source models.
A little over a year ago, a Chinese lab released a model called DeepSeek R1, surprising the AI industry. NVIDIA’s stock dropped 17% in a single day. It was called AI’s sputnik moment.
Not long ago, there was a large gap in intelligence between open-source models and closed-source models. When DeepSeek R1 launched, it considerably narrowed this gap.
DeepSeek R1 showed that for the first time, frontier AI wasn’t just in the hands of a few well-funded American labs. It proved that open-source intelligence came much closer to closed-source intelligence than previously thought possible. Today, there are well over 2 million open-source models on Hugging Face, which is 25x more than just 5 years ago.
There are over 2 million open-source models on HuggingFace. From Inference Engineering by Philip Kiely. Open-source models are now released frequently, but outside the AI space, many people don’t know much about them. What does it mean for a model to be open-source? When should you use an open-source model? Will open-source models ever catch up to closed-source models? And what does any of this mean for software engineers?
Here’s what you need to know.
What are open-source models?
When someone says "open-source models," they generally mean open-weight models. There is some debate over whether open-weight models are fully analogous to open-source. Traditional open-source fully exposes all source code for anyone to see or modify, whereas open-weight models only expose the weights, not necessarily the code or the training data.
Despite this dispute, open-weight models are still commonly referred to as open-source models.
To understand what it means for a model’s weights to be public, let’s briefly define a model’s weights.
What are model weights?
In simple terms, a weight is just a modifier on a variable that influences some outcome. In a million-parameter model, there are a million such modifiers, each one recorded as a weight. Update the right ones, and you shift the output. Want the model to talk like a pirate? You'd fine-tune it on pirate-speech examples, which updates weights broadly across the model until pirate-speak becomes the more likely output.
Open-source models typically refer to models with publicly available weights. Examples include models like DeepSeek, GPT-OSS, and GLM. The majority of open-source models are released under the MIT license, though some can restrict commercial use.
Open-weight models vs closed-weight models.Open-source models vs closed-source models
Closed-source models are models with weights that aren’t available to the public, such as GPT-5 or Claude Sonnet. With both of these closed-source models you can access them via APIs, but the training data and weights aren’t available to you.
Companies that offer closed-source models typically monetize them by bundling the model and the underlying costs of serving it (compute, for example) into a single per-output price. They typically charge per token. Some companies, like OpenAI and Google, offer both closed-source models and open-source models.
With an open-source model, the weights are available to the public. This allows any person or company to host the model. However, running a model can require serious GPU hardware, and this is where inference providers come in. Inference providers take open-source models, run them on their own infrastructure with multiple GPUs, optimize them using various techniques, and expose the models through a simple API to users. Inference providers typically charge by GPU time for custom (dedicated) deployments, or per token for pre-packaged model APIs.
This means that open- and closed-source models look mostly similar from a developer’s perspective. You hit an endpoint and get a response. The big difference is that with open-source models, the underlying model is open-weight, which means you have more choices about which provider to use, what you want to optimize for, and you have the ability to run the model yourself if you need to.
What can you do with open-source models?
Because open-source models allow you to view and modify their weights, there are several major benefits to using them.
Specialization
With a closed-source model, you’re working with a general-purpose system. You can prompt, but you can’t change the underlying model. What you can do with open-source models that you can’t do with closed-source ones is specialize them. Because you have access to the model’s weights, you can fine-tune the model with your own data, whereas with closed-source models like Claude or ChatGPT, the only lever you have is the prompt you provide.
Terms like “fine-tuning” and “post-training” refer to variations of the same core idea: taking a pretrained open model and specializing it. The result is a specialized model that performs well at a specific use case.
Initially, it was believed that open-source models could accomplish only a fraction of the tasks that their closed-source counterparts could. However, the specialization of open-source models has led to a significant increase in the number of use cases accomplished by open-source models.
Control
Open-source models allow you to optimize for your use case. As Baseten CTO Amir Haghighat points out, closed-source model APIs are designed to serve millions of users at high throughput, often at the expense of latency.
Latency is often critical, especially with AI voices or AI phone calls. When you deploy an open-source model, you have fine-grained control over the tradeoff between latency and throughput. With closed-source models, you don’t have this level of control.
Cost
A chart from Artificial Analysis comparing the cost of leading AI models per 1M tokens (3:1 input-output ratio), showing GPT OSS-120B as the cheapest at $0.3 and GPT-5.5 as the most expensive at $11.3 — highlighting the wide gap in pricing across today’s top LLMs.Open-source models average 87% lower cost than closed-source ones. This is due to a few factors: paying per GPU usage versus per token can be cheaper, competition between inference providers drives down prices, and massive amounts of research on model optimizations that make models more efficient.
When should (and shouldn’t) you use open-source models?
Baseten Head of Training Charlie O’Neill believes that most engineers building with AI don’t need to start out with an open-source model. In other words, start with the best closed-source model available and work your way down. See which closed-source models can handle your task.
Once you’ve stress-tested your task against closed models, do you have a baseline understanding to evaluate an open-source alternative? If you’re still searching for product-market fit, going with a closed-source model is a good default option.
Open-source models should be considered once your usage grows to the point where token costs on closed models become a concern. They’re also worth exploring if you have strict latency or throughput constraints.
Will open-source models ever catch up to closed-source ones?
Open-source models have been consistently 3-6 months behind closed models. An initial question you might ask is: will they catch up to closed-source ones?
But this may be the wrong question.
This is a point made in Inference Engineering: the important question isn’t whether open-source models are as good as the best closed-source models in general, but whether they are good enough for specific tasks. While open-weight models follow a few months behind the capabilities frontier pushed by closed-weight models, that lag can still leave them above the threshold needed for many real-world AI products. And with the right fine-tuning, open-source models can in some cases outperform closed ones on specific tasks.
Charlie O'Neill agrees that open-source models don’t necessarily need to beat frontier labs on benchmarks — they just need to cover economically valuable work. And much of this work requires specialization (fine-tuning on data for your specific use case), in which case open weights can be especially valuable.
The bigger questions on open-source models
At the heart of the open-source versus closed-source debate are deeper questions about AI. As AI gets more intelligent, who should have access to it?
Without open-source models, AI development would be concentrated in a handful of well-funded American labs. Open-source models distribute that power. This distribution means more developers, more innovation, and more access to AI. However, it also means less control over who uses it and how.
There’s also the question of compute. Even if many models are open-source, the compute required to train them at scale is not. Finally, there are important geopolitical considerations: what does the rise of open-source AI mean for the United States as Chinese labs increasingly lead in this domain?
These aren't questions with clean answers. But they're quietly shaping the future of AI, and we would be remiss not to acknowledge them. Over the next four parts, we'll get more in-depth on how open-source models actually work, when to use them, and what they mean for software engineers today.
