DeepSeek's publicly disclosed training cost is around $6 million, and the Llama series has accumulated hundreds of millions of downloads — the big-model companies package their "most powerful models" for free download. On the surface it looks like charity; in reality, it's a business bet measured in the tens of billions. This week, a beginner on Reddit asked "why are all the local AIs free," and our editorial team thinks this question deserves a minute of every non-technical reader's time.

What this is

The so-called "local AI" refers to running open-source models on your own computer or company server, not connected to paid APIs like OpenAI or Anthropic. This is possible because companies including Meta (Llama), Mistral, Alibaba's Qwen, and DeepSeek publish their trained model parameters — think of them as the model's "brain-wiring numbers" — as free downloads.

This is close to "open-source software," but the more precise term is "open weights": you get the parameters free, but the training data and training code typically aren't public. So "free" saves you the licensing fee, not all the costs — running locally usually requires an Nvidia GPU that costs tens of thousands of dollars.

Industry view

Supporters call this an open play: free models attract global developers to try and build on them. Once the ecosystem scales, cloud services, enterprise customization, and model licensing are where the real money is. Both Mistral and DeepSeek have raised billions on this logic. Meta's reasoning is even more direct — its core business is ads and cloud; AI is the gateway.

But the objections are equally clear. Security researchers point out that open weights drop the "gatekeeper," and downstream abuse is nearly impossible to trace. Business analysts suspect most open-source companies still haven't found a stable profit path, with valuations resting largely on expectations of "future earnings." There's a subtler concern too: when anyone can download the most powerful models, how do regulators enforce rules, and who bears the responsibility — no one has an answer yet.

Impact on regular people

  • For enterprise IT: The technical bar for building in-house models is dropping fast. Companies with healthy budgets are starting to keep sensitive data on-premises, but hardware and operations costs still keep small and medium businesses on the outside.
  • For individual careers: People who can tune open-source models to handle proprietary data will see their resumes become increasingly valuable — but a sizable gap still sits between "can deploy" and "can use well."
  • For consumer markets: Phone and home-appliance makers are experimenting with built-in small open-source models. Over the next year or two, offline AI features will become increasingly common.