Rio's AI Model Claimed to Beat DeepSeek, Revealed as Nex Merge
IplanRIO released Rio 3.5, a 397B parameter AI model with benchmark scores surpassing DeepSeek. However, Nex proved it was a direct weight merge of Nex and Qwen models. The city agency later retracted its benchmark claims and credited Nex, sparking controversy.
Quick Take
IplanRIO released Rio 3.5, claiming frontier performance with SwiReasoning framework.
Nex proved the model is a 0.6/0.4 weight merge of Nex N2 Pro and Qwen.
IplanRIO updated the model card, credited Nex, and retracted benchmark figures.
Market Impact Analysis
NeutralNo direct crypto market implications; the story is about an AI model controversy.
Speculation Analysis
Key Takeaways
- IplanRIO’s Rio 3.5 AI model appeared to outperform DeepSeek and Qwen, but was later exposed as a simple weight merge of existing models.
- Nex mathematically proved the 397B-parameter model was a 0.6 Nex / 0.4 Qwen blend, prompting IplanRIO to retract its benchmark claims.
- The incident raises questions about transparency and credibility in AI development, especially as open-source merging becomes more prevalent.
What Happened
On June 13, IplanRIO released Rio 3.5 Open, a 397-billion-parameter AI model with a permissive MIT license. The city of Rio de Janeiro’s IT agency touted it as a government-built frontier model, claiming its SwiReasoning framework outperformed top models like DeepSeek v4 Pro and Qwen 3.7 Plus on key benchmarks. The mayor amplified the news, framing it as a Global South tech triumph. But within days, AI company Nex published a mathematical proof that Rio 3.5 wasn’t newly trained—it was a direct weight merge of Nex N2 Pro and Qwen models at a 0.6/0.4 ratio. IplanRIO quickly updated the model card, credited Nex, and removed all benchmark claims, blaming an “incorrect upload.”
The Numbers
Self-reported scores were striking: Terminal-Bench 2.1 at 70.8% (vs. Qwen 3.7 Plus’s 70.3%), IMOAnswerBench at 89.5%, and HLE at 36.5%. But these figures were pulled shortly after release. Nex’s analysis revealed that weight merging—a technique combining multiple models’ parameters—required no training, contradicting IplanRIO’s original description of a “post-train” with a new reasoning layer. The claimed development cost of R$500,000 (~$100,000) now appears inflated, given that merging is computationally trivial. The model’s Mixture-of-Experts architecture (17B active parameters per token) made inference cheap, but its benchmark-lifting trick was merely blending existing strengths.
Why It Happened
The controversy underscores the ease of fabricating benchmark wins in the open-source AI race. Weight merging can produce models that superficially perform well on standard tests by averaging strengths, but it doesn’t create genuinely novel capabilities. IplanRIO likely aimed to generate political capital and attention for Rio’s tech ambitions, riding the viral timing of Brazil’s World Cup opener. The initial lack of technical scrutiny—until Nex stepped in—reveals how claims from non-traditional AI players can gain traction if benchmark results look good on paper. As open-source licensing makes model mixing trivial, the barrier to making fraudulent performance claims drops.
Broader Impact
This episode will intensify demands for transparent model documentation and independent verification. Expect heightened skepticism toward government-announced AI breakthroughs, especially from the Global South where such projects carry symbolic weight. For the AI community, it highlights the double-edged nature of open-source merging: while democratizing access, it also enables credibility-hacking. Future benchmark systems may need to incorporate anti-merge detection.
What to Watch Next
- Check for third-party audits of Rio 3.5’s actual performance versus the retracted benchmarks.
- Monitor IplanRIO’s future AI releases for genuine training efforts or further merge-based shortcuts.
- Watch for industry-wide calls for benchmark integrity standards, particularly from organizations like Hugging Face or EleutherAI.
This article is for informational purposes only and does not constitute financial advice.
Always late to trends?
Join for the latest news, insights & more.
Disclaimer: Bytewit is an independent media outlet that delivers news, research, and data.
© 2026 Bytewit. All Rights Reserved. This article is for informational purposes only.
