Rio de Janeiro's 'homegrown' 397B LLM allegedly copied from existing models
Original: Rio de Janeiro's "homegrown" LLM appears to be a merge of an existing model
Why This Matters
Highlights concerns about AI model authenticity and proper attribution in open-source development
Rio-3.5-Open-397B, presented as an original 397B parameter model by IplanRIO, is accused of being a direct merge of Nex-N2 and Qwen models without training. Evidence shows 0.6/0.4 weight blending across all layers.
Developer 00INDEX filed a GitHub issue claiming Rio de Janeiro's Rio-3.5-Open-397B model is not original but a mathematical merge of existing models. The accusation states the 397B parameter model combines 60% Nex-N2_pro weights with 40% Qwen3.5-397B weights without additional training. Evidence includes the deployed model identifying itself as 'Nex, from Nex-AGI' 79% of the time when system prompts are removed, and statistical analysis showing weight tensors match the alleged 0.6/0.4 blend 'to thousands of standard deviations' across all 60 network layers. The model was presented by IplanRIO as their original work.