sicutdeux@blog:~/links$cat rio-de-janeiro-s-homegrown-llm-appears-to-be-a-merge-of-an-existing-model.md
Rio de Janeiro's "homegrown" LLM appears to be a merge of an existing model
---
source_url:
source_name:
github.com
published:
2026-06-23
status:
published
---
Rio-3.5-Open-397B's weights are a direct element-wise merge of Nex with Qwen3.5-397B-A17B at roughly 0.6/0.4, with no evidence of independent training. Every weight tensor across all 60 layers matches this blend to thousands of standard deviations.
straightforward model forensics. the authors demonstrate weight interpolation through statistical analysis and behavioral testing—when you strip the system prompt, the model identifies itself as Nex. worth reading as a case study in reproducibility and the technical liability of claiming novel models without actual training work. the methodology (tensor-level analysis + identity testing) is pragmatic and replicable. relevant context: model merging is cheap, training is not, and attribution matters when claiming openness.