Curious Companion: Weight, What? An Introduction to Open Source AI Models

LISTEN TO THE EPISODE

In this episode I dig into open weight and open source AI models: what they actually are, why they keep showing up in AI news, and why I think they’re central to future-proofing our AI usage. I cover a bit of the technical aspect, the economic factors at play, and make the case that open weight models are a true hedge against getting stuck forever paying whatever a handful of AI companies decide to charge while simultaneously worrying that access could just be stripped away at any moment.

AI Agency

Not every episode is for everybody and some topics are a bit more niche than others. Today we’re talking about open source and open weight AI models, and while this topic doesn’t necessarily have a “try this, automate this, learn this” nature to it, I think it’s a super important topic as it relates to the future of AI and our ability to maintain some sort of agency and not be completely and totally subject to the whims of Anthropic, OpenAI, and the rest of those companies that literally don’t give two shits about us.

Why Open Models Are Even a Thing At All

So before I get into what open models actually are, let’s talk about why they’re even a thing at all; why they come up in AI conversation, why you’ll read about them in AI news, and why AI companies are publicly fighting about them (more on that at the end of this newsletter).

It really comes down to two things: how they affect the market, and what they can do for the individual user.

Now, remember, an LLM is just a computer program, albeit an insanely complex one. LLMs like Claude and ChatGPT are closed models, which means only Anthropic and OpenAI, respectively, can run that computer program. We just get to use the model. Open weight models on the other hand, give users, whether that’s companies or individuals, the ability to actually host and run these programs (the LLMs) themselves.

From a market standpoint, open weight models can help keep prices in check. Right now, a handful of companies (Anthropic, OpenAI, Google, etc) control access to their best models completely. Remember, those models are closed, meaning you can only use them, not own them. They set the prices, they can raise them whenever they want, and you either pay it or you don’t get to use their model.

Open models break that. Once a model is released, the recipe is out, permanently, and multiple companies can host and run that exact same model and then compete with each other on price.

Yes, there are companies that host and run open weight AI models and then charge per use, and because multiple of these hosting companies can all serve the same exact open model at the same time, they compete with each other on price, and that keeps the price down.

From an individual user standpoint, it’s really about sovereignty and not being dependent on one company’s decisions about access to a tool. If a closed model provider changes their pricing, changes their rules, or just shuts something down, you have zero say and zero access to that model. With an open model, you’re never fully at the mercy of one company, because the thing itself, the actual model, isn’t owned by a single gatekeeper..

In the most extreme, techy-heavy version of this scenario, you run the model locally, entirely on your own machine, and have basically full control over how you use it. Nobody can change your price, shut off your access, see your data, or take the model away from you. Depending on the size of the model you might need an insanely expensive computer setup, but everything would be…yours.

Open Weight vs. Open Source

Alright, so let’s get into definitions and learn a teeny bit about the tech side.

I’ve been using the term “open model,” but what you’ll often read about or see online are the terms “open weight” and “open source.”

Open weight = a company takes the model they trained and releases it, publishing the finished, trained version, that can be downloaded and run by an individual

Open source (the real, strict definition, and also a much bigger ask) = Developers publish the weights, AND also the training code and the actual training data, plus enough documentation that someone could rebuild the whole model from scratch if they wanted to. This is the actual bar for “open source,” and almost nothing out there actually clears it.

For My Nerds: Why It’s Called “Open Weight”

For my nerds in the crew, let’s briefly dive into why it’s called “open weight”. If this ain’t your jam feel free to jump ahead to the next section.

To set the foundation, let’s take it back to our super early episodes where I discussed parameters. LLM parameters can be a bit difficult to conceptualize, but in essence they are the connections between patterns in language (grammar, tone, rhythm, logic) with a weight assigned.

During training, the model gets fed a ton of content, basically everything on the internet, and then gets tested on that content. It gets presented parts of a real sentence, such as “The sky is ____,” and then gets asked what comes next. The model guesses (completes the sentence), and if it’s wrong, it uses math to adjust the weights between language patterns (the parameters) to increase the probability it will guess correctly the next time.

More parameters, aka more connections between patterns in language, means the more finely tuned responses from the LLM can be. This fine tuning occurs during “pre-training,” and then the weights are “frozen” (meaning they stop being tuned) prior to the model being deployed for public use.

So, an open weight model gives you those trained weights, the finished set of numbers that make the model work, ready to download and run. This is what’s being referenced whenever you hear the term “open weight model.”

The majority of what’s being released in the general “open model” category are actually open weight models, despite the fact that folks are using lazy language and also calling them open source.

“Open source AI” has basically been claimed by marketing at this point. Models you may have heard of, like Llama, DeepSeek, Qwen, Kimi, are open weight.

Genuinely open source models do exist, they’re just a completely different tier of the conversation, closer to academic infrastructure than a consumer product. These truly open source models are built by researchers who want other researchers to study exactly how a model got made, they’re not built by companies trying to win the “best model” race.

Who’s Making What

That list of names from the last section is a good segue into what open weight models are actually out there right now, and who’s putting them out:

Meta → Llama
Google → Gemma
Alibaba → Qwen
Moonshot AI → Kimi
DeepSeek → DeepSeek (model), e.g. DeepSeek-V4, DeepSeek-R1
Mistral AI → Mistral (model), e.g. Mistral Small, Mistral Large
Zhipu AI (also goes by Z.ai now) → GLM
OpenAI → gpt-oss

Hopefully after seeing names like Google and OpenAI on that list, you’re asking yourself why would these companies make these open weight models, and why would anyone actually choose to use them?

Great question 😉

Open Doesn’t Mean Free

Two things worth highlighting before we chat about the drivers of development and usage:

Open weight does not mean free, but it is typically associated with a lower price
In order for any of these models to make any kind of a splash, they need to be capable

If an open weight model is making headlines, it’s because it satisfies both of these criteria and is comparable to the frontier models (Claude, ChatGPT, etc), and it’s cheaper.

Regarding price, there are really three separate price points depending on how you’re accessing an open weight model. Run it fully locally, on your own hardware, and there’s no per-use charge at all. Your only cost is electricity, and whatever you dished out for the machine you’re running it on.

Access the open weight model through a hosting company or an API, and you’re paying per token, with published rates that tend to stay competitive precisely because multiple companies can host the same model and undercut each other.

Lastly, there’s the consumer app that many of the models have. Nearly every open model has a free tier, and it’s often generous, unlimited basic chat in some cases, with only the heavier features gated behind a paid plan. Again, the real pricing safety net here is that because the model itself is open, multiple companies can host it, which means they’re all competing with each other on price.

Regarding capability, so as to not bore you with benchmarks that neither of us understand and make this episode any longer than it needs to be: when it comes to capability, closed frontier models are still at the top of the leaderboard, BUT the open models are genuinely competitive.

The gap between the closed models and the best open weight models (namely Kimi K3) has gotten small enough that for the vast majority of what regular people and businesses actually do, that gap barely matters anymore, and in a few specific areas, open models are winning outright.

But Why?

So why would any of these companies release these open weight models?

Infiltration. If everyone builds their tools and startups on top of Llama, Meta doesn’t need to charge for it, they just made themselves the foundation everyone else depends on.
It’s free R&D. The second the weights are public, thousands of outside developers start poking at it, breaking it, fixing it, building tools on top of it, all without the company paying a single one of them.
It’s a recruiting move. The best AI researchers want their work seen and used, not buried in a closed product. Companies that publish get first pick of talent.
Undercut the market and attack the leader. If you’re not in first place, DeepSeek, Moonshot, Meta, undercutting the whole game by giving your version away for free is how you steal attention from whoever is.
The model itself isn’t the business. Meta makes money on ads. Google and Microsoft make money on the cloud. If a free model gets more people building on their infrastructure, that’s the actual payoff, the model’s just the hook.

I think it’s pretty clear why an individual user might opt for an open weight model, but allow me to list a few points that support their usage so as to not give all the attention to the companies and their motives for building these tools:

Cost. Multiple companies can host the same open model and compete on price, so you’re not stuck with whatever number one company decides to charge.
No lock-in. If your provider jacks up prices or disappears, you switch to a different host running the same model.
Privacy. Sensitive client work, medical stuff, legal stuff; you can run an open model entirely on your own machine and nothing ever leaves your computer. Closed models always send your data to their servers.
Customization. If you’re running the model yourself, whether that’s locally or on rented infrastructure, you can fine-tune an open model on your own data, your own voice, your own specific use case. With Claude or ChatGPT you’re limited to writing a really good prompt, you’re not touching what’s underneath.

How to Actually Try One

For those of you who are curious and want to give one of these open weight models a try, you’ve got options:

The free app or website. This is the simplest option, with no download, no setup, and it works exactly like using Claude. Most major open models have one: Kimi has its own app, DeepSeek has one, Google’s Gemma is available through AI Studio, Mistral has Le Chat. This is absolutely the right starting point for most people.
A hosted “playground” or API, no download required. If you want more control, or you’re a developer building something, you can access the model through the company’s API instead of their consumer app. This still doesn’t involve downloading anything, and you’re using their servers, just through a different door, usually billed per use instead of a flat monthly price. You can also go through a third-party host instead of the original company. Services like OpenRouter, Together AI, or Groq all host these same open models and let you access them, sometimes cheaper or faster than going direct.
Download it and run it yourself. This is the fully local option. You get the actual weights file from something like Hugging Face, and use a free tool like Ollama or LM Studio to run the model on your own computer. This only works for the smaller models like Llama’s 8-billion-parameter version, Gemma’s 4B or 12B sizes, or Qwen’s smaller 7B and 8B models, not the massive flagship versions making headlines. The largest open models, like Kimi K3, require far more computing power than a personal computer has.

The Nvidia Letter

I’ll probably give Kimi a try at some point, but more than looking for a new toy to try, I really just wanted to create this episode to better understand open weight models actually are, because I do think they have a ton of promise.

Fun fact: I’m not the only one who thinks this.

On July 24th, 2026, Nvidia CEO Jensen Huang publicly posted a letter co-signed by 25 companies, including Microsoft, Meta, and Hugging Face, arguing that America’s AI advantage depends on a broad, open ecosystem, not just one company’s best model.

This letter came in response to both accusations from the White House (insert massive eye roll) that Moonshot, a Chinese AI company, distilled Fable 5 in order to build their own model, and a bigger fight in Washington as to whether open models should be restricted outright.

Anthropic did NOT sign the letter, but their CEO did eventually put out a statement saying they’re not against open models as a category, they just want safety testing required across the board, open or closed. Meanwhile, the letter kept growing, with the signatory list doubling to fifty within a single day, with OpenAI and Google both joining shortly after.

It’s not lost on me that the letter was written by the CEO of Nvidia, a company that stands to benefit tremendously from anything AI related, open or closed. But I’ve expressed my concerns numerous times about the fact that companies like Anthropic and OpenAI have us by the proverbial short hairs, and I do believe that open weight models are an excellent solution.

How I Used AI This Week

Each episode I share a quick example of how I used AI that week.

This week I used a new (to me) Claude feature: Scheduled Tasks.

ChatGPT has had this feature, and I actually mentioned on a previous episode way back when I used it to set up a recurring search for deals on an outdoor camera and an SSD.

This time I set up a scheduled task in Claude that will run every Wednesday morning and deliver me a Prompting Curiosity brief.

When I sit down to make these episodes, I often chat with Claude about topic ideas, so I decided to automate that process and have Claude just deliver me some topic ideas every Wednesday morning, along with two ideas for things I could build with Claude to automate a task in my business.

The Scheduled Task has run once, and I already had this open weight topic in mind for today’s episode so I didn’t need a suggestion, BUT I did like one of the builds that it offered up. I’m busy AF right now so I don’t have time to be tinkering, but I’ve saved it to give a shot on a rainy day.

Da Wrap-up

Open weight models like Kimi K3, DeepSeek, and Llama aren’t fully open source, but what matters is that we know the difference in terminology and more importantly that we know these models exist.

Big players are releasing open weight models for their own (likely not so altruistic) reasons, but the upshot for us is more competition, more control, and at the end of the day, a way to possibly future-proof our access to a tool that has proven to be incredibly useful to many of us.

As always, endlessly grateful for you and your curiosity.

Catch you next Thursday.

Maestro out.