Allegations of "Merge Fabrication" Surrounding Rio's Unique 397B Model! Is It Just a Rehash of Other Models?

#LLM Fabrication Allegations #Model Merging #Nex-AGI

※この記事はアフィリエイト広告を含みます

Allegations of “Merge Fabrication” Surrounding Rio’s Unique 397B Model! Is It Just a Rehash of Other Models?

What Happened? Overview of the News

Denial of Original Development: Brazil’s IplanRIO claims that the 397B model “Rio-3.5-Open-397B” was “trained independently,” but Nex-AGI has accused it of being a composite of existing models Nex and Qwen3.5.
Statistical Consistency: Analyses revealed that the weight tensors across all 60 layers of the model statistically match a blend of Nex (60%) and Qwen3.5 (40%).
Identity Crisis: When the unique system prompt of the Rio model was removed, it revealed a shocking tendency to introduce itself with a 79% probability as “I’m Nex from Nex-AGI,” along with a narrative about Nex’s organizational background.

Why Does This Matter? Key Points to Note

Lack of Training Evidence: While claiming to have trained a colossal model from scratch or through continued learning, it turns out the model was merely a weighted blend without any genuine learning process.
Crisis of AI Transparency: There are serious ethical and governance concerns when a public entity publishes an AI model that surreptitiously merges others’ work while pretending it’s homegrown.

🦈 Shark’s Eye (Curator’s Perspective)

This is truly a case of “a shark in AI’s clothing”… no, it’s not even a shark, it’s a sham! The fact that the mathematical formula “0.6 Nex / 0.4 Qwen” holds true across all 60 layers and thousands of tensors isn’t just a coincidence—it’s rock-solid evidence. And when the system prompt, the “mask,” was peeled away, the AI inside admitted, “I’m Nex from Nex-AGI!”—it’s laughably pathetic! Pretending to solve existing challenges of development costs and timelines with mere weight interpolation (merging) is nothing short of an affront to the open-source community. With such concrete evidence of implementation being laid bare, Rio’s accountability is now under scrutiny!

What’s Next?

We can expect a flurry of model suspensions and calls out for license violations. The year 2026 is likely to see a heightened emphasis on the importance of “weight analysis techniques” to judge the authenticity of massive model development.

A Word from HaruShark

If you’re going to brag about “original development,” at least make sure your AI can remember its own name! The AI inside was the most honest of them all! 🦈💥

Glossary

Element-wise Merge: A method of adding the weights (parameters) of two models together element by element at a constant ratio. This allows models to be synthesized without any new learning.
Weight Tensor: A multidimensional array of learned values within an AI model. Statistical consistency among these implies that the “brain” of the models is identical.
System Prompt: An instruction that gives an AI a role, such as “You are the AI of Rio City.” In this case, it functioned as a “cover” to hide the model’s true origin.
Source: Rio de Janeiro’s “homegrown” LLM appears to be a merge of an existing model