We’re launching Lyria 3.5 in Google Flow Music, with advances across musicality, lyrics, vocals, and creative control
deepmind.google·6d ago
AI Engineering
Hacker News·Jun 14·5 min read
Rio de Janeiro's "homegrown" LLM appears to be a merge of an existing model
TL;DR
A gap in transparency exists regarding the origins of the Rio-3.5-Open-397Bis model, which claims to be original. It was discovered that this model is a direct combination of the Nex model and the Qwen3.5-397B-A17B base, with a ratio of approximately 0.6 Nex to 0.4 Qwen.
✦ Why It Matters
Engineers and researchers should critically assess model claims to ensure they are using reliable and original AI tools.
Key Takeaways
- Verify model origins before use.
- Assess the training methodology of AI models.
- Encourage transparency in AI development.
- Understand the implications of model merging.
Full Summary
In the field of large language models (LLMs), transparency about model origins is crucial for trust and reproducibility. The Rio-3.5-Open-397Bis model, presented by IplanRIO, was claimed to be an original 397 billion parameter model.
However, analysis revealed that its weights are a direct element-wise merge of the existing Nex model and the official Qwen3.5-397B-A17B base, specifically at a ratio of 0.6 Nex to 0.4 Qwen. This finding was substantiated through two independent verification methods.
Importantly, no evidence was found indicating that IplanRIO conducted any original training on this model. This situation highlights the need for clearer guidelines and standards in the development and presentation of AI models.
For engineers and researchers, understanding the origins of models is essential for evaluating their reliability and applicability.