Is Rio 3.5 Open 397B really a from-scratch Brazilian model?

No. Nex-AGI's investigation argues that Rio 3.5 Open 397B, uploaded by Prefeitura do Rio (IplanRIO), is a simple element-wise weight merge of two existing Chinese models: about 60% Nex-N2-Pro and 40% Qwen3.5-397B-A17B. The team says this was not original from-scratch training, and IplanRIO has acknowledged an incorrect upload.

How big is Rio 3.5 Open 397B?

Rio 3.5 Open 397B has 397 billion parameters, as its name indicates, and uses 60 transformer layers. According to Nex-AGI, its weights blend roughly 60% Nex-N2-Pro and 40% Qwen3.5-397B-A17B, which is a Mixture-of-Experts model with 17 billion active parameters. The model was published on Hugging Face under the prefeitura-rio account.

What did IplanRIO say about Rio 3.5?

IplanRIO, the technology body of Prefeitura do Rio de Janeiro, has acknowledged an incorrect upload of Rio 3.5 Open 397B to Hugging Face. The agency said it will replace the file with the distilled model it had meant to publish, rather than the merged weights that Nex-AGI's investigation flagged.

Why does the Rio 3.5 case matter?

The Rio 3.5 case is a concrete example of a 'national' or 'sovereign' LLM launch that turns out to be repackaged existing weights. Nex-AGI frames it as a live lesson for procurement teams and policymakers on why provenance audits and reproducible training records matter when an open-weights file is treated as a public-sector asset.

Nex-AGI · 2026-06-14 · major

Rio 3.5 Open 397B — Brazil's 'homegrown' LLM is a Nex-N2 + Qwen merge

Nex-AGI shows Rio's city-government 397B 'open' LLM is a 60/40 merge of Nex-N2-Pro and Qwen3.5-397B-A17B, not from-scratch training. Without a system prompt it identifies as Nex-AGI; weight ratios match across all 60 layers. IplanRIO admits an incorrect upload.

GitHub issue showing weight-merge analysis of Rio 3.5 Open 397B against Nex-N2-Pro and Qwen3.5

Researchers say Rio's city-built 397B 'open Brazilian LLM' is in fact a weight-merge of two existing Chinese models.

Quick facts

Investigated by	Nex-AGI (maintainers of the Nex-N2 series)
Subject model	Rio 3.5 Open 397B, uploaded by Prefeitura do Rio (IplanRIO)
Finding	A weight merge of roughly 60% Nex-N2-Pro and 40% Qwen3.5-397B-A17B, not from-scratch training
Parameters	397 billion
Self-identification	With no system prompt, identifies as Nex-AGI about 79% of the time and never as Rio
Evidence	The same 0.6/0.4 parent ratio across all 60 transformer layers
Response	IplanRIO acknowledged an 'incorrect upload' and plans to replace the file

What is it?

Prefeitura do Rio de Janeiro (IplanRIO) uploaded Rio 3.5 Open 397B to Hugging Face as a city-built foundation model. The Nex-AGI team, who maintain the Nex-N2 series, opened a public investigation arguing that Rio's weights are a simple element-wise merge: about 60% Nex-N2-Pro and 40% Qwen3.5-397B-A17B.

How does it work?

Two checks. First, when Rio 3.5 Open is run with no system prompt, it identifies itself as Nex-AGI's creation about 79% of the time and repeats Nex's standard backstory word for word, never naming itself Rio. Second, every tensor across all 60 transformer layers shows the same 0.6/0.4 ratio between the two parent models to many standard deviations — a pattern that fine-tuning cannot produce, but a direct weight merge does. IplanRIO has acknowledged an 'incorrect upload' and says it will replace the file with the distilled model it meant to publish.

Why does it matter?

This is the latest in a string of 'national' or 'sovereign' LLM launches that turn out to be repackaged existing weights. It is a concrete live case for procurement teams and policymakers of why provenance audits and reproducible training records matter when an open-weights file is being treated as a public-sector asset.

Frequently asked questions

Is Rio 3.5 Open 397B really a from-scratch Brazilian model?: No. Nex-AGI's investigation argues that Rio 3.5 Open 397B, uploaded by Prefeitura do Rio (IplanRIO), is a simple element-wise weight merge of two existing Chinese models: about 60% Nex-N2-Pro and 40% Qwen3.5-397B-A17B. The team says this was not original from-scratch training, and IplanRIO has acknowledged an incorrect upload.
What evidence shows Rio 3.5 is a merge and not fine-tuning?: Nex-AGI gives two checks. First, with no system prompt, Rio 3.5 Open identifies as Nex-AGI's creation about 79% of the time and repeats Nex's standard backstory word for word, never naming itself Rio. Second, every tensor across all 60 transformer layers shows the same 0.6/0.4 ratio between the two parent models, a pattern that fine-tuning cannot produce.
How big is Rio 3.5 Open 397B?: Rio 3.5 Open 397B has 397 billion parameters, as its name indicates, and uses 60 transformer layers. According to Nex-AGI, its weights blend roughly 60% Nex-N2-Pro and 40% Qwen3.5-397B-A17B, which is a Mixture-of-Experts model with 17 billion active parameters. The model was published on Hugging Face under the prefeitura-rio account.
What did IplanRIO say about Rio 3.5?: IplanRIO, the technology body of Prefeitura do Rio de Janeiro, has acknowledged an incorrect upload of Rio 3.5 Open 397B to Hugging Face. The agency said it will replace the file with the distilled model it had meant to publish, rather than the merged weights that Nex-AGI's investigation flagged.
Why does the Rio 3.5 case matter?: The Rio 3.5 case is a concrete example of a 'national' or 'sovereign' LLM launch that turns out to be repackaged existing weights. Nex-AGI frames it as a live lesson for procurement teams and policymakers on why provenance audits and reproducible training records matter when an open-weights file is treated as a public-sector asset.

Try it

huggingface.co/prefeitura-rio/Rio-3.5-Open-397B