Most people know that when it comes to large language models (LLMs), not all are created equal. Some models are clearly more “intelligent,” while others offer benefits like being significantly cheaper to use.
Yet all LLMs fall into one of three categories: proprietary, open weight, or open source. Understanding these distinctions is the key to understanding the benefits and drawbacks of any individual LLM.
Proprietary LLMs
Let’s start with the first category, proprietary LLMs. As its name suggests, proprietary LLMs are owned by a single entity, usually a large corporation. Even if you don’t know the names of any open source or open weight LLMs, you almost certainly know the proprietary ones.
These include the biggest names in AI, such as OpenAI’s GPT and o-series that powers ChatGPT (GPT-5.6 and o3-mini, for example), Anthropic’s Claude, and Google’s Gemini.
Not only are proprietary LLMs privately owned by some of the largest tech giants, but they are also incredibly closed-off and opaque. Their code and training data are highly protected and largely a mystery to those outside the company, and if you want to use them, you need to do so through the cloud.
Of course, there are benefits to these closed-off proprietary models. Most are considered “frontier” models because they are seen as being the most intelligent and capable LLMs out there. This is largely due to the computational power behind them and the sheer volume of data they were trained on. But all that costs massive amounts of money, which is why frontier models are mostly owned by major companies.
These proprietary models also have another significant benefit: For the most part, they are incredibly simple for businesses to adopt thanks to their APIs, which allow businesses to easily “plug” their workflows into the LLMs.
Open Weight LLMs
The second category of LLMs is known as open weight models. These are frequently confused with open source models (more on those below), but in fact they are distinct.
As noted by LM-Kit, in LLM parlance, a “weight” can be likened to a large language model’s memory. An “open weight” LLM means that the LLM’s creator lets people download that large language model’s memory.
Perhaps the best-known open weight model is Meta’s Llama (at least in the U.S.). But while Llama is the exception among American LLMs, which are mostly proprietary, elsewhere in the world it’s an entirely different story.
In China, the most well-known and widely used LLMs are nearly all open weight models. These include many that have made headlines for their capabilities and low cost, such as DeepSeek, Moonshot’s Kimi, and Alibaba’s Qwen.
For businesses looking to adopt an LLM into their workflow, open weights can be very appealing since they can be significantly cheaper than proprietary LLMs when it comes to cost per token and are increasingly nearly as capable as the best frontier models. They also offer the option of more privacy.
This is because open weight models can usually be downloaded from the provider’s cloud and run locally on a computer or a business’s private server. If an LLM is run locally, a business can ensure that any information it feeds to the LLM (such as business plans, contracts, or other sensitive information) stays in the business’s hands.
But one big drawback to open weight LLMs compared to proprietary ones is that they aren’t “plug and play” if run locally. When a business uses a proprietary cloud-based LLM like Claude, it simply plugs into Claude’s APIs and lets Anthropic’s servers do the heavy lifting. When a company deploys an open weight LLM on its own servers, it needs to pay for the computers it runs on. In other words, though the price per token for an open weight LLM can be much lower, open weight LLMs can carry significant upfront hardware costs for businesses if run locally.
Open Source LLMs
Finally, we get to open source LLMs. Though many people mistake open weight LLMs for open source, they are different. Open weights, while not as inscrutable as proprietary models, are still relatively opaque: the businesses that choose to use them generally don’t have access to the datasets the models were trained on.
But with open source LLMs, it’s different. “Open source” is a general, decades-old term for software that can be downloaded, verified, modified, and distributed freely, and it is now being used to describe LLMs that meet those criteria.
As noted by The Open Source Initiative (OSI), a nonprofit, open source models are fully transparent. Not only can users download the models to run locally on their machines, but their full training code and training datasets (or at least data sources, if the datasets aren’t legally permitted to be copied) are also available for examination. Open source models can also be modified and shared completely, something that is not possible with proprietary models and, for open weights, only possible for the weights (in a process known as “fine-tuning“).
The malleability and transparency of open source LLMs are potential advantages for businesses in a world where regulators are increasingly taking a more hardened stance towards the AI systems organizations run.
“Policymakers need concrete proof that AI models comply with laws on privacy, discrimination, and consumer protection,” notes OSI. “With Open Weights, regulators see only the end result, not the steps taken to reach it. Full openness eases the burden of proving a model’s compliance across various jurisdictions.”
According to the OSI, LLMs that meet the definition of open source include AI2’s OLMo, Eleuther AI’s Pythia, Google’s T5, and LLM360’s Amber and CrystalCoder.
Which type of LLM is best?
Figuring out which type of LLM is “best” largely depends on your definition of “best.”
If you mean the most capable models with the largest, constantly updated training datasets, a proprietary frontier model like GPT-5.6, Claude, or Gemini can arguably be considered the best.
But if you mean “best return on investment,” it is hard not to conclude that open weight models like DeepSeek and Kimi may be the best choice if run locally, especially after upfront costs like hardware have been accounted for
—or even if run on the open weight company’s cloud. This is because their capabilities now nearly match those of the most prominent frontier models, and they are also much less expensive when it comes to price per token, the chunks of data an LLM processes.
For example, according to July 2026 data from Spheron, OpenAI’s flagship GPT-5.6 Sol and Anthropic’s Claude Opus 5 / Opus 4.8 cost $5 per million input tokens, while Google’s Gemini 3.1 Pro cost $2 per million input tokens (up to 200,000; then the price jumps to $4 per million input tokens). Contrast that with the most well-known open weight models, such as DeepSeek; it costs just $0.435 per million input tokens for its V4-Pro model.
And if your definition of “best” is a model that you have the utmost control over, and with training data you can verify, there is arguably no better option than an open source model, which, also, of course, has the benefit of being free (minus the ongoing cost of compute, of course).
