"Open source AI" gets thrown around loosely, and the reality is messier than the label suggests โ some "open" models release only their weights, others release training code too, and the licenses attached to each vary a lot in what you're actually allowed to do. Still, a real and important category has emerged: models you can download, run on your own hardware or cloud account, fine-tune, and deploy without depending on a single vendor's API. Llama, Mistral, and DeepSeek are the three names that come up most often in this conversation, and they represent three different approaches to the same basic idea.
What "Open Weights" Actually Means
Most of what gets called open source AI is more precisely "open weights" โ the trained parameters of the model are published for anyone to download, but the training data, and often the exact training code, stay private. This distinction matters because it affects what you can actually verify about a model: you can run it, test it, and fine-tune it, but you generally can't reproduce it from scratch or fully audit what it was trained on. That's a meaningfully different kind of openness than, say, an open source programming language, but it's still a huge practical shift from closed, API-only models, because it means the model itself โ not just access to it โ belongs to whoever downloads it.
The license attached to the weights matters just as much as the weights themselves. Some releases allow essentially unrestricted commercial use; others carry conditions, like requiring large companies above a certain size to negotiate a separate license, or restricting use in specific ways. Anyone deploying an open model in a commercial product should actually read the license rather than assume "open" means "unrestricted."
Llama: Meta's Widely Adopted Family
Meta's Llama models are probably the most widely used open-weight family, both because Meta has released multiple generations at a range of sizes and because the ecosystem of tools built around Llama โ fine-tuning scripts, quantized versions, hosting options โ is unusually mature. Llama models come in different sizes, from versions small enough to run on a decent consumer GPU or even a laptop, up to much larger versions that need serious server hardware.
That range of sizes is a big part of the appeal: a small Llama model can run entirely on your own machine with no data leaving your device, which matters for privacy-sensitive use cases, while the larger versions compete more directly with proprietary models on general capability. If you want to compare a Llama-based experience against a closed alternative, Llama vs Mistral walks through how the two open families differ in practice, and the same kind of comparison applies to the newer Chinese entrant discussed below.
Mistral: The European Challenger
Mistral is a French AI lab that built its reputation on releasing capable models that punch above their parameter count โ meaning a comparatively smaller, more efficient model that performs competitively against larger ones, which matters a lot for the cost and speed of actually running the thing. Mistral has released both open-weight models and, alongside them, hosted API products and a consumer-facing chat product, Mistral Le Chat, which is worth trying if you want a sense of the underlying model's capabilities without setting up your own infrastructure.
Mistral's positioning has also mattered for reasons beyond pure capability: as a European company, it's frequently discussed in the context of data sovereignty and reducing reliance on U.S.-based AI providers, which is a real consideration for European businesses and public sector organizations even when it's not the loudest part of the technical conversation.
DeepSeek: Efficiency as the Headline
DeepSeek is a Chinese AI lab that gained significant attention for releasing open-weight models that achieved strong performance while reportedly requiring less computational resources to train than comparable Western models โ a claim that shook up assumptions about how much raw compute spending was actually necessary to build a frontier-competitive model. Whatever the exact numbers turn out to be, the broader effect was real: DeepSeek's releases pushed the entire field to take efficiency-focused training more seriously, and demonstrated that capable open models weren't the exclusive domain of the best-funded U.S. labs.
DeepSeek offers both a chat product and downloadable model weights, and it's worth trying directly if you want to judge its output quality for yourself rather than relying on secondhand claims. As with any model, especially one developed under a different regulatory and political environment, it's worth being thoughtful about what data you send to a hosted version versus running the open weights locally, particularly for sensitive business use.
Why Anyone Chooses Open Over Closed
There are a handful of recurring, legitimate reasons teams choose an open-weight model over a closed API:
Cost at scale. If you're making millions of API calls, running your own instance of an open model โ especially a smaller, efficient one โ can be dramatically cheaper than paying per-token to a hosted provider, assuming you already have or are willing to manage the infrastructure.
Data control. Running a model on your own servers means your data never leaves your environment, which matters a lot for regulated industries or anyone with strict data residency requirements.
Customization. Open weights can be fine-tuned on your own data in ways that hosted APIs typically don't allow, letting you build a model specialized for a narrow task โ customer support for one specific product line, for instance โ that outperforms a general-purpose model on that narrow task.
No vendor lock-in. If a provider changes its pricing, deprecates a model, or shifts its terms of service, an organization depending entirely on that one API has no fallback. Running open weights, or at least having the option to, is a hedge against that.
Where to Actually Run These Models
For individuals and small teams who want to run open models locally without managing infrastructure themselves, Ollama has become one of the most popular ways to do it โ it packages open-weight models, including Llama and Mistral variants, into a simple local runtime that works on a regular laptop. Hugging Face is the other essential piece of this ecosystem: it hosts the actual model files for thousands of open-weight models, along with the tools and documentation needed to fine-tune and deploy them, functioning as something close to the GitHub of the open model world. For a comparison of local-first versus hosted-hub approaches, see Ollama vs Hugging Face.
For production deployment at scale, hosted inference providers like Together AI and Groq run open-weight models on optimized infrastructure so you get the openness and cost benefits of the underlying model without managing GPUs yourself. OpenRouter takes a slightly different angle, acting as a single API that routes to many different underlying models โ open and closed โ so you can switch between them without rewriting your integration each time.
Quality, Not Just Cost
It's worth being clear-eyed about capability too. The gap between the best open-weight models and the best closed models has narrowed considerably, but "narrowed" isn't the same as "closed entirely" โ the most capable proprietary models still tend to lead on the hardest reasoning tasks, the largest context windows, and the most polished tool-use behavior, simply because their developers can pour enormous, sustained resources into a single flagship release. Open models, meanwhile, often win on a different axis: how good a model is per dollar of compute spent running it, rather than raw peak capability regardless of cost. For many real business tasks โ summarizing documents, classifying support tickets, drafting routine content โ that per-dollar efficiency matters more than winning on the hardest possible benchmark, which is exactly why open models have found such a strong foothold in production systems even without leading the capability race outright.
It's also worth remembering that "open" doesn't mean "static." Mistral, Meta, and DeepSeek all continue to release updated versions of their model families, and the gap in either direction shifts with each release cycle. Treating this as a fixed hierarchy โ closed models are always smarter, open models are always cheaper โ will get out of date quickly; treating it as an evolving trade-off you re-evaluate periodically is a more durable way to think about it.
The Practical Trade-off
None of this means open models are simply "better" โ closed, proprietary models from providers like OpenAI and Anthropic generally remain at or near the frontier of raw capability, and they come with none of the infrastructure burden. The honest way to think about it is a trade-off: closed models offer convenience and typically the highest capability ceiling; open models offer control, customizability, and cost efficiency at scale, at the price of needing someone on your team who can actually manage the deployment. Many organizations end up using both โ a closed frontier model for the hardest problems, and a fine-tuned open model for high-volume, narrower tasks where the cost difference adds up fast. Understanding that trade-off, rather than treating "open source" as an automatic virtue, is what actually helps you pick the right tool. For more on how this fits into the broader shift toward AI that does things rather than just answers questions, see our guide to AI coding agents, several of which โ including Aider and Continue โ are explicitly built to work with open models like these.
