Allegations of Plagiarism: Is Rio's "Homegrown" LLM Just a Model Merge?

A controversy has erupted within the open-source AI community following a detailed report posted to the nex-agi/Nex-N2 GitHub repository. The issue, opened by user 00INDEX on June 14, 2026, suggests that the model Rio-3.5-Open-397B—which was marketed as an original creation developed by IplanRIO—is not an independent work at all.

The Core Accusation

According to the report, the Rio-3.5-Open-397B model is actually a direct, element-wise merge of two existing models. The claimants argue that there is zero evidence of any proprietary training conducted by the Rio team.

The "Recipe" for Rio

The mathematical composition of the model is alleged to be:

$\text{Rio-3.5-Open-397B} \approx (0.6 \times \text{Nex-N2\_pro}) + (0.4 \times \text{Qwen3.5-397B-A17B})$

Below is a conceptual visualization of how this merge was likely constructed:

Evidence Provided

The authors of the GitHub issue present two distinct methods of verification to prove their claim:

1. Behavioral Analysis (The "Identity Crisis")

When the hard-coded system prompt (which instructs the model: "You are Rio") is stripped away, the model's true identity emerges.

"Its own deployed model identifies itself as 'Nex, from Nex-AGI' 79% of the time — and as 'Rio' 0% of the time."

Identity Distribution Table:

Identity Claimed	Frequency	Status
~~Rio~~	0%	False
Nex (Nex-AGI)	79%	True
Other/Unknown	21%	N/A

Furthermore, the model reportedly recites the exact bespoke backstory of the Nex-AGI organization, word-for-word.

2. Technical Weight Analysis

The most damning evidence comes from the model's architecture. The researchers analyzed the weight tensors across the entire network.

Scope: All 60 layers of the model.
Finding: Every single component of the network matches the $0.6/0.4$ blend.
Precision: The match is consistent to thousands of standard deviations.

To illustrate this in a simplified programmatic sense, the merge looks like this:

# Conceptual representation of the alleged merge
def merge_models(model_nex, model_qwen):
    rio_weights = []
    for w_nex, w_qwen in zip(model_nex.weights, model_qwen.weights):
        # Element-wise weighted average
        merged_layer = (0.6 * w_nex) + (0.4 * w_qwen)
        rio_weights.append(merged_layer)
    return rio_weights

Summary of Findings

The following checklist summarizes the discrepancies between the claims made by IplanRIO and the findings by Nex-AGI:

Claim: Original 397B model trained by IplanRIO. $\rightarrow$ Reality: Merge of Nex and Qwen.
Claim: Unique identity as "Rio". $\rightarrow$ Reality: Identifies as "Nex" without system prompts.
Claim: Independent weights. $\rightarrow$ Reality: Mathematical match to $0.6/0.4$ ratio across 60 layers.

Involved Parties

The report was filed by the following user:

User: @00INDEX

As of the latest update, the issue remains open on GitHub, and the community awaits a response from the developers of the Rio model.