A GitHub issue on the Nex-N2 repository sent the HN underground into a frenzy this week. The crux of the matter: Rio de Janeiro apparently launched what was marketed as a "homemade" large language model — a kind of technological pride symbol for the city. The problem? A handful of community members dug into the model weights and architecture, and believe it looks suspiciously like a merge of already-existing open-source models.

For those not in the loop: a "merge" means blending the weights from two or more existing models together — it's a legitimate technique, but it's not the same as training a model from scratch. It's roughly the difference between baking your own bread versus buying two loaves and mixing the slices.

Convincing people you've built an LLM from scratch is easy — actually doing it costs between $50,000 and $100 million.

And that's exactly where the shoe pinches. Research into what it actually costs to train an LLM from the ground up makes for brutal reading: even a small 7B model can cost hundreds of thousands of dollars in compute alone, and a proper frontier model can easily top $100 million. That makes it easy for a trained community eye to smell something off when a city suddenly announces its "own" model without comparable resources.

Rio fooled everyone: The city's 'homegrown' AI is just a remix - Bilde 1

What makes this HN thread especially interesting isn't just the revelation itself — it's the speed. The community investigation apparently didn't take long. Model merging is a widespread practice in the open-source world (and certainly not illegal if the licenses permit it), but marketing the result as a "homegrown" innovation is another matter entirely. It creates false expectations and obscures proper credit for those who actually trained the original models.

This pattern isn't new. We've seen similar cases before where companies or institutions take existing Llama- or Mistral-based models, do a bit of fine-tuning or merging, and present the result as something revolutionary. But with an increasingly sophisticated open-source community, getting away with it is becoming harder.

Source assessment: This is based on community findings from HN and GitHub. No independent technical audit has been published yet, and the Nex-N2 / Rio side has not issued a public response as of now. Stay tuned — this one could develop quickly.