Rio-3.5-Open-397B: Analysis of the 'Homegrown' LLM Merge Controversy

The Rio-3.5-Open-397B large language model (LLM), initially presented as a homegrown fine-tune of Qwen3.5, was discovered by the community to be a mathematical merge of two existing open-weight models. This incident highlights the ongoing challenges regarding model provenance and transparency in the open-source AI ecosystem.

The Discovery of the Model Merge

Technical analysis of the model weights revealed that Rio-3.5-Open-397B is not a traditional fine-tune, but a weighted interpolation of existing models. Specifically, the model's weight tensors are a 60/40 blend of Nex-N2 Pro and Qwen3.5-397B-A17B across all 60 layers of the network.

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.

This discovery was brought to light via a GitHub issue, where researchers noted that Nex-N2 had been released approximately one week prior to the Rio model, making the timing of the merge highly suspicious.

Official Response and Correction

Following the public scrutiny, the municipality of Rio de Janeiro (via its IT company, IplanRIO) updated the model's Hugging Face page to acknowledge the merge. The updated documentation now explicitly states that the model is a merge of Nex-N2-Pro and Qwen3.5-397B-A17B, followed by on-policy distillation from a stronger model.

The developers apologized for the confusion, claiming that the "base merged version" was uploaded by mistake instead of the final distilled model. They stated:

The model is built via a merge of https://huggingface.co/nex-agi/Nex-N2-Pro and https://huggingface.co/Qwen/Qwen3.5-397B-A17B , proceeded by On-Policy Distillation from a stronger model. We detected an incorrect upload in the previous version, where the base merged version was upload instead of the final distilled model.

Technical Implications of Model Merging

Model merging involves the mathematical combination of weights from two or more models that share the same architecture. Unlike distillation or fine-tuning, which require significant compute for training, a simple linear combination of weights can sometimes enhance model performance without degrading it.

In the case of Rio-3.5-Open-397B, the community observed that a simple linear combination of weights did not degrade performance, which serves as a testament to the robustness of current deep learning architectures. This process allows developers to combine the capabilities of different models without the need for extensive new training data or compute resources.

Community Reaction and Ethical Concerns

The incident sparked significant debate within the AI community regarding the ethics of rebranding merged models as "homegrown" and the lack of transparency in model provenance.

Lack of Transparency

Critics argue that presenting a merge as a proprietary or homegrown effort undermines trust in the open-source community. The lack of disclosure regarding the use of Nex-N2 Pro and Qwen3.5-397B-A17B was seen as a deceptive practice.

Provenance Tracking

There is a growing call for better provenance tracking and transparency standards for model releases. The community noted that this pattern of rebranding merged models is not isolated, but a recurring issue in the LLM space, where some developers claim original work while profiting from the work of others without proper attribution.