Rio de Janeiro's "homegrown" LLM appears to be a merge of an existing model
A new claim alleges that 'Rio-3.5-Open-397B', an LLM presented by Rio de Janeiro as a homegrown creation, is in fact a direct merge of existing models. The exposé provides two strong, independent lines of evidence to support this claim, stirring questions about the originality and transparency in AI development. This technical detective work resonates with the HN community's interest in uncovering the true origins and methodologies behind prominent AI projects.
The Lowdown
A recently published report asserts that the 'Rio-3.5-Open-397B' language model, promoted by IplanRIO on behalf of the city of Rio de Janeiro as an original creation, is not genuinely homegrown. Instead, the authors claim it is an element-wise merge, approximately 60% Nex and 40% Qwen, with no discernible evidence of independent training by IplanRIO. The report offers compelling technical evidence to back this assertion.
- Self-Identification with Removed Prompt: When Rio's hard-coded 'You are Rio' system prompt is removed, the deployed model identifies itself as 'Nex, from Nex-AGI' 79% of the time, and never as 'Rio'. It even reproduces Nex-AGI's unique backstory verbatim.
- Weight Tensor Analysis: Every single weight tensor across all 60 layers and components of the Rio model is demonstrably a 0.6/0.4 blend of Nex and Qwen, with statistical significance 'to thousands of standard deviations'. This specific interpolation cannot be explained by other fine-tuning methods.
The findings challenge the official narrative of Rio-3.5-Open-397B as an original model, presenting it instead as a highly specific blend of pre-existing, publicly available models.