TL;DR
Get business pricing on tech for your team
- Business-only prices and quantity discounts
- Tax-exempt purchasing
- Multiple users, one account, clear invoices
Rio de Janeiro’s local large language model appears to be a direct merge of existing models Nex and Qwen, not a new, independently trained system. This revelation questions the model’s claimed originality and training process.
Recent analysis by Hacker News suggests that Rio de Janeiro’s claimed ‘homegrown’ large language model (LLM) is not an original training effort but a direct blend of existing models Nex and Qwen, raising questions about its purported independence and development process.
The model, named Rio-3.5-Open-397B and presented as an original 397-billion-parameter LLM trained by IplanRIO, has been shown to be a weighted combination of two publicly available models. Hacker News researchers found that, when the model’s system prompt is removed, it identifies itself as ‘Nex, from Nex-AGI’ 79% of the time, and as ‘Rio’ only 0%.
Further analysis revealed that every weight tensor in the Rio model matches a 0.6/0.4 blend of Nex and Qwen models across all layers and components, with no evidence of additional training or fine-tuning. This suggests the model is essentially an interpolation of existing models rather than a new, independently trained system.
These findings challenge the claims made by IplanRIO regarding the model’s originality, indicating it may be a simple merge rather than a novel development.
Implications for Model Authenticity and Development Claims
This discovery impacts perceptions of local AI development efforts, raising concerns over transparency and originality. If the model is a blend rather than a trained system, it questions the credibility of claims made by IplanRIO and could influence future trust in regional AI projects. It also highlights broader issues around transparency in model training and the importance of clear disclosures in AI development.
Products Worth Considering
Apple MacBook Air 15-inch with M5 chip, 24GB RAM, 512GB SSD, Midnight
Samsung Galaxy Tab A7 Lite 8.7 32GB Wi‑Fi Android Tablet – Long‑Lasting Battery, Compact Slim Design, Sturdy Metal Frame (US Version)
Sony WH-1000XM4C (M4 2nd Gen) Wireless Noise-Cancelling Headphones, Black
As an affiliate, we earn on qualifying purchases.
Background on Rio’s AI Development Claims
Rio de Janeiro announced the development of an ‘original’ 397B parameter LLM, branded as Rio-3.5-Open-397B, purportedly trained by local efforts through IplanRIO. The model was presented as a homegrown solution aimed at boosting regional AI capabilities. However, prior to this analysis, little was publicly known about its training process or origin.
Recent scrutiny by Hacker News researchers reveals that the model’s weights are a direct merge of two existing models, Nex and Qwen, with no signs of independent training. The model’s self-identification behavior and tensor analysis support this conclusion, casting doubt on the originality claims.
“Every weight tensor in Rio is, to thousands of standard deviations, the same 0.6/0.4 blend of Nex and Qwen across all layers.”
— Hacker News researcher
Extent of Training and Fine-tuning Still Unclear
It remains unclear whether any additional fine-tuning or modifications were applied beyond the weight merging. The analysis suggests no evidence of training, but this has not been definitively confirmed by IplanRIO or other sources.
Further Investigation and Official Clarifications Expected
IplanRIO or affiliated authorities are likely to respond with official statements addressing these findings. Additional technical analysis or disclosures may clarify whether any further training or fine-tuning occurred. The broader AI community will also scrutinize the implications of this discovery for regional AI development claims.
Key Questions
Is Rio de Janeiro’s LLM truly a homegrown model?
Based on recent analysis, the model appears to be a weighted merge of existing models Nex and Qwen, not an independently trained system.
What evidence supports the claim that the model is a merge?
Tensor analysis shows every weight tensor matches a blend of Nex and Qwen models, and the model identifies itself as Nex most of the time when prompted differently.
Could the model have been fine-tuned after merging?
It is currently unclear; the analysis found no signs of additional training, but official confirmation is pending.
Why does this matter for regional AI development?
If the model is not truly original, it raises questions about transparency and claims of local innovation in AI technology.
What are the next steps for this story?
Official statements from IplanRIO are expected, along with further technical analysis to verify training details and assess the impact on regional AI efforts.
Source: Hacker News
