Controversy Over Rio 3.5 397B Model: Allegations of Mismanaged Funds in LLM Development

A complex situation is unfolding in the Large Language Model (LLM) landscape, raising questions about transparency and integrity in the development of funded projects. The model in question, named Rio 3.5 397B, is at the center of serious allegations of fraud and improper fund management. These implications extend beyond the specific project, affecting the trust of the community and investors in the artificial intelligence sector. This incident highlights the need for rigorous due diligence and clear accountability, crucial aspects for organizations evaluating LLM deployments, especially in self-hosted or on-premise contexts where control and data sovereignty are paramount.

The Timeline of Events and Technical Discrepancies

The controversy began with funding of approximately R$500,000 (equivalent to about $100,000 USD) allocated for the training of the Rio 3.5 397B model. The project's initial documentation claimed that the model was developed based on Qwen 3.5 397B, incorporating advanced training processes and significant improvements. However, subsequent investigations revealed a very different reality: the uploaded model was, in fact, a simple merge with Nex N2 Pro, without any substantial additional training.

This technical discrepancy is fundamental. LLM training involves a computationally intensive process, requiring significant hardware resources (such as GPUs with high VRAM) and considerable energy expenditure to process vast datasets and refine the model's capabilities. A simple merge, conversely, combines the architectures or weights of existing models. While useful in certain contexts, this operation does not equate to new training and does not justify the same costs or development claims. Following the discovery, the model's documentation was updated to admit the Nex N2 Pro base, while still insisting on additional training and attributing the error to uploading the wrong version. The previously available model was then removed from platforms like Hugging Face. In an attempt at damage control, the team later communicated via social media that the final trained model was "lost," necessitating a restart of development from scratch.

Implications for Trust and Data Sovereignty

This incident raises serious concerns for the entire LLM ecosystem. For companies and institutions considering the adoption of AI solutions, particularly those opting for on-premise or air-gapped deployments for reasons of data sovereignty and compliance, trust in a model's declared provenance and capabilities is crucial. Incidents like that of Rio 3.5 397B undermine this trust, highlighting the risks associated with unverified claims and the potential misappropriation of funds intended for research and development projects.

The Total Cost of Ownership (TCO) of an LLM is not limited to hardware or energy costs; it also includes the cost of due diligence, verification, and risk mitigation. If a model is presented under false pretenses, investment decisions in infrastructure, personnel, and integration can be severely compromised. Transparency thus becomes not only an ethical but also a strategic requirement, especially in a sector where technical complexity makes it difficult for non-specialists to discern the truth behind claims.

Future Outlook and the Need for Verification

The developer community and investors are called to greater vigilance. The availability of Open Source platforms and the ease of deploying pre-trained models can sometimes mask a lack of real added value or significant training. This case underscores the importance of independent benchmarks and rigorous peer review to validate model performance and origins.

For those evaluating on-premise deployments, AI-RADAR offers analytical frameworks on /llm-onpremise to assess the trade-offs between control, performance, and costs. The lesson from Rio 3.5 397B is clear: the promise of innovation must be supported by verifiable facts and impeccable ethical conduct. Only then can an AI future built on trust and genuine technological progression be achieved.