# Rio 3.5

> **TL;DR.** Rio 3.

- **Category:** AI / Open Source / Models
- **Stage:** nascent
- **Age:** 3 days
- **Origin date:** 2026-06-13
- **First detected:** 2026-06-16
- **Canonical URL:** https://earlyterms.com/term/rio-3-5
- **Sources:** 7 primary URLs

## Definition

Rio 3.5 Open 397B is a 397-billion-parameter open-weight language model released June 13, 2026 by [IplanRIO](https://huggingface.co/prefeitura-rio/Rio-3.5-Open-397B), the municipal IT company of Rio de Janeiro — the first frontier-class AI model shipped by a city government.

The release became AI news for the wrong reason: within 24 hours, Nex-AGI proved via tensor-level weight analysis that Rio 3.5 was a 0.6/0.4 element-wise merge of their model Nex N2 Pro and the base Qwen 3.5-397B-A17B, not original development. IplanRIO apologized, credited Nex, and retracted the benchmark claims — leaving a landmark case study in AI provenance fraud.

## Analogy

A city printing its name on a Ferrari that's really two other cars welded together.

## Why it's emerging now

On June 13, 2026, IplanRIO released Rio 3.5 Open 397B claiming it outperformed DeepSeek and Qwen — a first for any municipal government. By June 14, Nex-AGI had mathematically proven the model was an undisclosed merge of Nex N2 Pro and Qwen weights. IplanRIO apologized and retracted benchmark claims, producing AI's sharpest recent case of provenance misrepresentation.

## Related terms

- *parent:* Qwen 3.5
- *related:* qwen3
- *related:* Nex N2 Pro
- *related:* deepseek-v4
- *parent:* model merge
- *related:* grpo
- *child:* SwiReasoning
- *related:* benchmark gaming
- *parent:* open source LLM
- *related:* ai-slop
- *related:* IplanRIO

## Sources

1. [IplanRIO — Rio 3.5 Open 397B model card on Hugging Face (updated post-controversy)](https://huggingface.co/prefeitura-rio/Rio-3.5-Open-397B)
2. [Decrypt — Rio de Janeiro Built an AI Model That Beat DeepSeek — But Was Based on Someone Else's Work](https://decrypt.co/371210/rio-ai-model-beat-deepseek-ownership-dispute-nex)
3. [Nex-AGI — Forensic weight-merge analysis proving Rio-3.5 = 0.6 Nex + 0.4 Qwen](https://github.com/nex-agi/Nex-N2/issues/4)
4. [Squared Tech — Rio LLM Exposed: tensor-level proof of model merge](https://www.squaredtech.co/rios-official-ai-model-is-proven-to-be-a-model-merge)
5. [Kingy AI — Rio 3.5 Open 397B: specs, benchmarks, and independent analysis](https://kingy.ai/news/rio-3-5-open-397b-specs-benchmarks-analysis/)
6. [Hacker News — Rio de Janeiro city model beats Qwen3.7 (142 points, 43 comments)](https://news.ycombinator.com/item?id=48527634)
7. [FelloAI — Rio 3.5 Open 397B technical overview](https://felloai.com/rio-3-5-open-397b/)

---
_Generated by EarlyTerms · https://earlyterms.com/term/rio-3-5_