Most individuals know that in terms of massive language fashions (LLMs), not all are created equal. Some fashions are clearly extra “clever,” whereas others supply advantages like being considerably cheaper to make use of.
But all LLMs fall into considered one of three classes: proprietary, open weight, or open supply. Understanding these distinctions is the important thing to understanding the advantages and downsides of any particular person LLM.
Proprietary LLMs
Let’s begin with the primary class, proprietary LLMs. As its title suggests, proprietary LLMs are owned by a single entity, often a big company. Even for those who don’t know the names of any open supply or open weight LLMs, you nearly actually know the proprietary ones.
These embrace the most important names in AI, resembling OpenAI’s GPT and o-series that powers ChatGPT (GPT-5.6 and o3-mini, for instance), Anthropic’s Claude, and Google’s Gemini.
Not solely are proprietary LLMs privately owned by a number of the largest tech giants, however they’re additionally extremely closed-off and opaque. Their code and coaching knowledge are extremely protected and largely a thriller to these exterior the corporate, and if you wish to use them, you want to take action by way of the cloud.
After all, there are advantages to those closed-off proprietary fashions. Most are thought of “frontier” fashions as a result of they’re seen as being essentially the most clever and succesful LLMs on the market. That is largely because of the computational energy behind them and the sheer quantity of information they have been educated on. However all that prices large quantities of cash, which is why frontier fashions are principally owned by main firms.
These proprietary fashions even have one other important profit: For essentially the most half, they’re extremely easy for companies to undertake because of their APIs, which permit companies to simply “plug” their workflows into the LLMs.
Open Weight LLMs
The second class of LLMs is named open weight fashions. These are steadily confused with open supply fashions (extra on these beneath), however the truth is they’re distinct.
As noted by LM-Package, in LLM parlance, a “weight” will be likened to a big language mannequin’s reminiscence. An “open weight” LLM signifies that the LLM’s creator lets individuals obtain that enormous language mannequin’s reminiscence.
Maybe the best-known open weight mannequin is Meta’s Llama (a minimum of within the U.S.). However whereas Llama is the exception amongst American LLMs, that are principally proprietary, elsewhere on the planet it’s a completely completely different story.
In China, essentially the most well-known and extensively used LLMs are practically all open weight fashions. These embrace many who have made headlines for his or her capabilities and low price, resembling DeepSeek, Moonshot’s Kimi, and Alibaba’s Qwen.
For companies seeking to undertake an LLM into their workflow, open weights will be very interesting since they are often considerably cheaper than proprietary LLMs in terms of price per token and are more and more practically as succesful as the perfect frontier fashions. In addition they supply the choice of extra privateness.
It is because open weight fashions can often be downloaded from the supplier’s cloud and run domestically on a pc or a enterprise’s non-public server. If an LLM is run domestically, a enterprise can be sure that any data it feeds to the LLM (resembling enterprise plans, contracts, or different delicate data) stays within the enterprise’s fingers.
However one massive downside to open weight LLMs in comparison with proprietary ones is that they aren’t “plug and play” if run domestically. When a enterprise makes use of a proprietary cloud-based LLM like Claude, it merely plugs into Claude’s APIs and lets Anthropic’s servers do the heavy lifting. When an organization deploys an open weight LLM by itself servers, it must pay for the computer systems it runs on. In different phrases, although the worth per token for an open weight LLM will be a lot decrease, open weight LLMs can carry important upfront {hardware} prices for companies if run domestically.
Open Supply LLMs
Lastly, we get to open supply LLMs. Although many individuals mistake open weight LLMs for open supply, they’re completely different. Open weights, whereas not as inscrutable as proprietary fashions, are nonetheless comparatively opaque: the companies that select to make use of them typically don’t have entry to the datasets the fashions have been educated on.
However with open supply LLMs, it’s completely different. “Open supply” is a common, decades-old time period for software program that may be downloaded, verified, modified, and distributed freely, and it’s now getting used to explain LLMs that meet these standards.
As noted by The Open Supply Initiative (OSI), a nonprofit, open supply fashions are absolutely clear. Not solely can customers obtain the fashions to run domestically on their machines, however their full coaching code and coaching datasets (or a minimum of knowledge sources, if the datasets aren’t legally permitted to be copied) are additionally obtainable for examination. Open supply fashions may also be modified and shared fully, one thing that’s not potential with proprietary fashions and, for open weights, solely potential for the weights (in a course of often known as “fine-tuning“).
The malleability and transparency of open supply LLMs are potential benefits for companies in a world the place regulators are more and more taking a extra hardened stance in direction of the AI methods organizations run.
“Policymakers want concrete proof that AI fashions adjust to legal guidelines on privateness, discrimination, and shopper safety,” notes OSI. “With Open Weights, regulators see solely the top end result, not the steps taken to succeed in it. Full openness eases the burden of proving a mannequin’s compliance throughout numerous jurisdictions.”
Based on the OSI, LLMs that meet the definition of open supply include AI2’s OLMo, Eleuther AI’s Pythia, Google’s T5, and LLM360’s Amber and CrystalCoder.
Which kind of LLM is greatest?
Determining which sort of LLM is “greatest” largely relies on your definition of “greatest.”
Should you imply essentially the most succesful fashions with the most important, consistently up to date coaching datasets, a proprietary frontier mannequin like GPT-5.6, Claude, or Gemini can arguably be thought of the perfect.
However for those who imply “greatest return on funding,” it’s onerous to not conclude that open weight fashions like DeepSeek and Kimi could also be the only option if run domestically, particularly after upfront prices like {hardware} have been accounted for
—or even when run on the open weight firm’s cloud. It is because their capabilities now nearly match these of essentially the most outstanding frontier fashions, and they’re additionally a lot inexpensive in terms of value per token, the chunks of information an LLM processes.
For instance, in response to July 2026 data from Spheron, OpenAI’s flagship GPT-5.6 Sol and Anthropic’s Claude Opus 5 / Opus 4.8 price $5 per million enter tokens, whereas Google’s Gemini 3.1 Professional price $2 per million enter tokens (as much as 200,000; then the worth jumps to $4 per million enter tokens). Distinction that with essentially the most well-known open weight fashions, resembling DeepSeek; it prices simply $0.435 per million enter tokens for its V4-Professional mannequin.
And in case your definition of “greatest” is a mannequin that you’ve got the utmost management over, and with coaching knowledge you possibly can confirm, there may be arguably no higher choice than an open supply mannequin, which, additionally, in fact, has the advantage of being free (minus the continuing price of compute, in fact).
