Rio de Janeiro’s municipal IT company IplanRIO presented Rio 3.5 Open 397B as an open AI model, trained with public funds and outperforming DeepSeek V4 Pro and Qwen 3.7 Plus across a number of benchmarks. However, a day later the team of AI developer Nex stated that the tool looks like a direct merge of Nex-N2-Pro and Qwen3.5-397B-A17B.

After the claims, IplanRIO updated the Rio 3.5 Open 397B card on Hugging Face. The new description states that the model was built through a merge of Nex-N2-Pro and Qwen3.5-397B-A17B with subsequent distillation from a stronger model.

How Rio 3.5 was presented

IplanRIO posted Rio 3.5 Open 397B on Hugging Face on June 13, 2026 under the MIT license. The original description called the project a "frontier-level" general-purpose AI system and stated that the model was fine-tuned on the basis of Qwen3.5-397B-A17B.

The specifications mentioned 397 billion parameters, of which 17 billion are activated when processing each token. This architecture is called Mixture-of-Experts (MoE): the model uses not all parameters at once, but only a portion of specialized blocks.

IplanRIO also claimed a context window of 1.01 million tokens and the use of SwiReasoning. In the project description, this framework is presented as a mechanism that switches the model between explicit and latent reasoning modes.

The first version of the project card cited test results according to which Rio 3.5 outperformed Qwen 3.7 Plus and DeepSeek V4 Pro. On Terminal-Bench 2.1 the model scored 70.8% versus 70.3% for Qwen 3.7 Plus and 67.9% for DeepSeek V4 Pro. On Humanity’s Last Exam the figure was 36.5% versus 34.7% for Qwen 3.7 Plus, and on IMOAnswerBench — 89.5%.

After the release, the mayor of Rio de Janeiro, Eduardo Cavalieri, wrote on X that an open AI model trained in Rio with public funds had "surpassed all other models."

What Nex stated

On June 14, Nex opened an issue in the Nex-N2 repository on GitHub. The company stated that Rio 3.5 Open 397B is presented as an original IplanRIO model, but its weights look like a direct element-wise merge of Nex-N2-Pro and Qwen3.5-397B-A17B.

According to Nex’s estimate, Rio 3.5 consists of approximately 60% Nex-N2-Pro and 40% Qwen3.5-397B-A17B. The company claims it found no signs of independent training by IplanRIO.

Nex gave two arguments. After removing the "You are Rio" system prompt, the model, according to the company, called itself "Nex, from Nex-AGI" in 79% of responses and never once called itself Rio. Nex also stated that every weight tensor of Rio repeats the 0.6/0.4 proportion between Nex and Qwen across all 60 layers of the model.

"There is no innocent explanation for this," the Nex statement says.

In a separate post, the firm put the claim more simply: Rio 3.5 is essentially Nex’s open-source N2 Pro model "in a different wrapper."

Why the benchmarks raised questions

Decrypt noted that Nex-N2-Pro shows higher results in its own tests than Rio 3.5 in the original card. The Nex-N2-Pro description on Hugging Face lists 75.3% on Terminal-Bench 2.1 versus 70.8% for Rio 3.5. On GDPval the Nex model scored 1585 points versus 1533 for Rio.

As the outlet noted, if Rio is indeed a mixture of Nex-N2-Pro and Qwen3.5-397B-A17B, then its weaker results compared with Nex look expected. At the same time, the Rio 3.5 benchmarks themselves were removed from the main description after the card was updated.

How IplanRIO responded

After the claims, IplanRIO changed the model’s README on Hugging Face. The current version states that Rio 3.5 Open 397B was built through a merge of Nex-N2-Pro and Qwen3.5-397B-A17B and then underwent distillation.

Distillation is a training method in which one model adopts the behavior of a stronger model. In this case, IplanRIO claims it was supposed to publish not the base version, but the final distilled model.

"We regret the confusion and apologize," the updated README says.

The team also reported that it is working on re-uploading the correct model. There was no separate detailed public comment from IplanRIO, apart from the updated README, at the time of publication.

What the dispute is about

Using open models is not in itself a violation. Nex-N2-Pro is published under the Apache 2.0 license, and Qwen3.5-397B-A17B is also available as an open model. Such licenses allow models to be used, modified and distributed subject to the terms.

The dispute arose over the presentation of Rio 3.5. The original card gave the impression of an independent development and fine-tuning on the basis of Qwen3.5-397B-A17B, but did not name Nex-N2-Pro as one of the sources. In the open-source community this is seen as a transparency problem. Merging open weights, fine-tuning and distillation have become common practice, but developers are expected to disclose the source models and the contribution of third-party teams.

Earlier, Alibaba presented a family of "hybrid" AI models, Qwen3, which are "able to match or in some cases surpass" the best solutions from Google and OpenAI.

As a reminder, the Chinese AI startup DeepSeek presented DeepSeek-R1 in January 2025. That model became one of the main events of the AI market at the time.

Source: ForkLog