Rio de Janeiro's 'homegrown' LLM appears to be a merge of an existing model

The city of Rio de Janeiro’s IT agency, IplanRIO, released a large language model last week that it billed as an original 397-billion-parameter model trained from scratch. Evidence posted to a GitHub issue on June 14 suggests otherwise: the model’s weights appear to be a direct element-wise merge of two existing open-source models, with no additional training.

The claim comes from 00INDEX, a developer at the AI startup Nex-AGI. In the issue, 00INDEX writes that Rio-3.5-Open-397B is “a direct element-wise merge of our model, Nex, with the official Qwen3.5-397B-A17B base” in a ratio of approximately 0.6 Nex to 0.4 Qwen. The post includes two independent lines of evidence: behavioral testing and mathematical weight analysis.

First, the behavioral test. With Rio’s hard-coded “You are Rio” system prompt removed, the model was asked to identify itself. According to the post, it identified as “Nex, from Nex-AGI” 79% of the time and as “Rio” 0% of the time. It also recited Nex-AGI’s “bespoke backstory word-for-word.”

Second, the weight analysis. 00INDEX states that “every weight tensor in Rio is, to thousands of standard deviations, the same 0.6/0.4 blend of Nex and Qwen” across all 60 layers and every component of the network. The post claims that “other finetunes cannot be explained as interpolations,” meaning the weight pattern is inconsistent with any amount of post-merge training.

The implication is that IplanRIO did not train a model. It merged two existing models and presented the result as original work.

Model merging is a legitimate and common technique in open-source AI research. Tools like MergeKit and frameworks such as TIES, DARE, and SLERP allow practitioners to combine the weights of multiple models to produce a new model with blended capabilities. The technique is widely used by hobbyists and researchers to create instruction-tuned variants or domain-specific models without the cost of full training.

But there is a difference between merging and training. A merge is a mathematical operation on existing weights. Training requires forward and backward passes, loss functions, and compute. The distinction matters for reproducibility, attribution, and scientific honesty.

IplanRIO has not yet responded to the claim. The GitHub repository for Rio-3.5-Open-397B remains public. The model card on Hugging Face describes it as “an original 397B model trained by IplanRIO.” If the evidence in the Nex-AGI issue is correct, that description is false.

This is not the first time a public-sector AI project has faced questions about originality. In 2024, India’s BharatGPT project was criticized for using a wrapper around existing models without clear attribution. In 2025, a European government-funded LLM was found to be a fine-tune of Llama 3 with minimal architectural changes. The pattern is consistent: organizations with limited compute and ML expertise announce “homegrown” models that are actually derivatives of existing work.

The difference here is the specificity of the evidence. Weight-level interpolation analysis is rare in these controversies. Most derivative models are caught through behavioral testing or architecture comparison. The Nex-AGI post provides mathematical proof that Rio-3.5-Open-397B is a linear combination of two known models, not a trained model.

For the AI research community, this story is a reminder that model provenance is increasingly verifiable. Weight watermarking, interpolation analysis, and behavioral fingerprinting are becoming standard tools. Labs and institutions that attempt to pass off merges as original training risk public exposure.

For policymakers and funders, the story raises a harder question. If a city government’s IT agency cannot or will not train an original model, should it be funding AI projects at all? Rio’s investment in this model is unclear, but the reputational cost is already visible. The model’s Hugging Face page shows 187 stars on the Nex-N2 repository, but the issue thread is the more relevant metric of community trust.

What happens next depends on IplanRIO. A correction, an updated model card, and a transparent explanation of the merge process would go a long way. Silence or a defensive response will confirm the worst interpretation.

The Nex-AGI developer’s closing line in the issue is worth noting: “Judge for yourself.” The evidence is public. The math is clear. The question is whether IplanRIO will acknowledge it.

The broader lesson for AI builders is that the era of plausible deniability in model provenance is ending. Weight-space analysis is cheap and definitive. If you did not train the model, do not say you did.