IplanRIO has introduced the open AI model Rio 3.5 - ForkLog

# IplanRIO introduced the open AI model Rio 3.5

The municipal IT company Rio de Janeiro IplanRIO presented Rio 3.5 Open 397B as an open AI model trained with government funds and surpassing DeepSeek V4 Pro and Qwen 3.7 Plus in several benchmarks. However, a day later, the AI development team Nex claimed that the tool appears to be 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 merging Nex-N2-Pro and Qwen3.5-397B-A17B with subsequent distillation from a stronger model.

How Rio 3.5 was presented

IplanRIO uploaded Rio 3.5 Open 397B to Hugging Face on June 13, 2026, under the MIT license. In the initial description, the project was called a "state-of-the-art" general-purpose AI system and indicated that the model was fine-tuned based on Qwen3.5-397B-A17B.

The specifications mentioned 397 billion parameters, of which 17 billion are activated per token processed. This architecture is called Mixture-of-Experts (MoE): the model does not use all parameters at once but only a subset 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 implicit reasoning modes.

The first version of the project card included test results showing Rio 3.5 outperforming Qwen 3.7 Plus and DeepSeek V4 Pro. On Terminal-Bench 2.1, the model scored 70.8% compared to 70.3% for Qwen 3.7 Plus and 67.9% for DeepSeek V4 Pro. In Humanity’s Last Exam, the score was 36.5% versus 34.7% for Qwen 3.7 Plus, and in IMOAnswerBench — 89.5%.

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

🇧🇷 Open-source AI model trained in Rio with public funding over the past year by @Prefeitura_Rio surpassing all other models. Artificial intelligence is not something distant, foreign, or a billion-dollar laboratory… it’s not just for text, images… https://t.co/GK1ThytVV9

— Eduardo Cavaliere (@CavaliereRio) June 14, 2026

What Nex claimed

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

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

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

"There is no innocent explanation for this," — Nex’s statement.

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

The Rio 3.5 model broke the internet this week. The plot twist? It’s essentially our open-source model, Nex N2 Pro, wearing a different hat.

🤯 We analyzed the weights, and the recipe is exact: Rio 3.5 ≈ 0.6 * Nex N2 Pro + 0.4 * Qwen 3.5

It even literally introduces itself… pic.twitter.com/yHRRu37aut

— Nex (@NexEcosystem) June 14, 2026

Why benchmarks raised questions

Decrypt noted that Nex-N2-Pro shows higher results in its own tests than Rio 3.5 in the initial card. On Hugging Face, Nex-N2-Pro is listed with 75.3% on Terminal-Bench 2.1 versus 70.8% for Rio 3.5. On GDPval, Nex scored 1585 points versus 1533 for Rio.

As the publication pointed out, if Rio is indeed a mixture of Nex-N2-Pro and Qwen3.5-397B-A17B, then its weaker results compared to Nex are expected. Meanwhile, the Rio 3.5 benchmarks 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 merging Nex-N2-Pro and Qwen3.5-397B-A17B, then distilled.

Distillation is a training method where one model adopts the behavior of a stronger model. In this case, IplanRIO claims it should have published not the base version, but the final distilled model.

"We regret the confusion and apologize," — the updated README states.

The team also said it is working on re-uploading the correct model. As of the publication, IplanRIO had not provided a separate detailed public comment beyond the updated README.

What is the core of the dispute

Using open-source models itself is not 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 use, modification, and distribution of models under certain conditions.

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

Earlier, Alibaba introduced the "hybrid" Qwen3 family of AI models, which "can match or sometimes surpass" the best solutions from Google and OpenAI.

Recall that the Chinese AI startup DeepSeek introduced DeepSeek-R1 in January 2025. This model became one of the major events in the AI market at that time.