Brazilian AI company Nex-AGI is facing scrutiny after its 'homegrown' large language model, Nex-N2-7B, was identified in a GitHub issue as a merge of existing open-source models, contradicting the company's claims of building it from scratch.
A 'Homegrown' Claim Under Scrutiny
Nex-AGI had presented Nex-N2-7B as evidence of independently developed, sovereign AI capability from Brazil. The company did not disclose details of its training data, architecture, or compute resources, which prompted members of the open-source community to examine the model's files directly.
The GitHub Investigation
According to the GitHub issue, the investigation found several signs that the model was not built from scratch:
- Identical architecture: The model's
config.jsonfile matched another open-source model, Emerald-7B-V2.1. - Matching file sizes: The model's
.safetensorsweight files were byte-for-byte identical in size to the suspected source models. - Tensor analysis: Examination of the model's tensors indicated it appears to be a direct merge of two open-source models, Emerald-7B-V2.1 and Wey-L-7B-V3.
Model merging is a legitimate and common technique for combining existing models into new ones with blended capabilities. But presenting a merged model as one built entirely from scratch would be a misrepresentation of its origins.
Why It Matters
The episode illustrates how open-source communities can independently verify or challenge claims made by AI companies through code-level scrutiny of released model files. As more companies market proprietary or 'sovereign' AI models, this kind of transparency check — comparing config files, weight sizes, and tensor structures against known open-source models — is likely to remain one of the few ways outside parties can verify how a model was actually built.