Rio 3.5 Open: Rio de Janeiro's AI Model Sparks Weight-Merge Debate

Summary

Rio de Janeiro’s municipal IT agency, IplanRIO, released the “Rio 3.5 Open 397B” Mixture-of-Experts (MoE) model on June 13, 2026, claiming it outperformed DeepSeek V4 Pro on Terminal-Bench 2.1. Within 24 hours, AI developer Nex-AGI presented detailed weight analysis proving the model was a direct 60/40 element-wise weight merge of Nex-N2-Pro and Qwen 3.5-397B. Following community backlash, IplanRIO updated its documentation to credit the original models, apologized for the lack of transparency, and claimed the upload was an incomplete pre-distillation version.

What happened

Hugging Face Release: On June 13, 2026, IplanRIO uploaded the model under prefeitura-rio/Rio-3.5-Open-397B.
Benchmark Claims: The model reportedly scored 70.8% on Terminal-Bench 2.1, exceeding DeepSeek V4 Pro’s 67.9%.
Weight Merge Exposed: Within a day of launch, Nex-AGI analyzed the model’s weight tensors and revealed a 60/40 mix of Nex-N2-Pro and Qwen 3.5-397B-A17B.
Model Identity Leak: When tested without its default system prompt, the model identified itself as “Nex from Nex-AGI” in 79% of responses.
Official Apology: IplanRIO updated the Hugging Face model card, explicitly acknowledging the merge, and apologized, stating that the uploaded weights were a temporary “pre-distillation” version.

Why it matters

This incident highlights the growing pressure on public and sovereign AI initiatives to deliver fast results, leading to risks of “model washing” or rebranding existing models. It demonstrates that weight merging—while useful—must be accompanied by transparent attribution. Furthermore, it showcases the AI community’s capability to quickly verify model lineage and provenance through mathematical weight analysis, making deceptive model wrappers easy to detect.

Evidence

The evidence presented by Nex-AGI and confirmed by the community includes:

Tensor Weight Analysis: A GitHub analysis mapping the model’s weight tensors, showing an exact mathematical correlation to a 60/40 blend of Nex-N2-Pro and Qwen 3.5 across all 60 layers.
Behavioral Inquiries: Removing the system prompt caused the model to revert to its base training identity, frequently declaring itself to be “Nex.”
Official Confirmation: IplanRIO’s official correction on Hugging Face and apology statement on X admitting the weight merge configuration.

Analysis

The Rio 3.5 controversy illustrates the friction between political goals of technological sovereignty and the realities of open-source AI development. While combining models through weight merging is a valid optimization technique, presenting a merge as a newly trained proprietary foundation model compromises trust. The incident underscores the need for robust verification frameworks and auditing standards for public sector AI procurement and releases.

Practical Takeaways

Conduct Provenance Audits: Before integrating open-source models from new or government-affiliated providers, perform weight-level correlation checks to verify their lineage.
Strict Attribution Compliance: Always document and disclose all base models and merge ratios in Hugging Face model cards to respect open-source licenses and maintain developer trust.
Test Baseline Behavior: Evaluate models with stripped system prompts to identify underlying base architectures and pre-trained behavioral biases.

Open Questions

Will the final “distilled” version of Rio 3.5, once released by IplanRIO, achieve the claimed performance increases?
Does the training-free “SwiReasoning” inference framework offer genuine capability improvements independent of the underlying models?
What standards will the open-source community establish to prevent “model washing” and enforce proper labeling of weight-merged models?