Why a Brazilian Municipal IT Company Just Open-Sourced a Frontier AI Model

The juxtaposition could not be more stark. On June 12, 2026, the U.S. government ordered Anthropic to block foreign nationals from its most advanced models, Claude Fable 5 and Mythos 5, forcing the company to take them fully offline . Two days later, a municipal IT company owned by the city government of Rio de Janeiro published a massive, 397-billion-parameter model on Hugging Face, free for anyone in the world to download under a permissive MIT license .

The model is called Rio 3.5 Open 397B, and its release is a landmark moment in the debate over AI access, sovereignty, and the decentralization of frontier capability. But separating the political symbolism from the technical reality is essential.

What Rio 3.5 Open 397B Actually Is

Let’s be clear about the origin story. This is not a model trained from the ground up by a scrappy government IT team. IplanRIO, the municipal technology company of Rio de Janeiro’s city government, created the model by post-training from Alibaba’s Qwen 3.5-397B-A17B, a powerful open-weight foundation model . The developers describe the process as a model merge followed by on-policy distillation from a stronger model .

Subsequent community analysis has suggested the merge is a straightforward combination: approximately 0.6 of the open-source model Nex N2 Pro and 0.4 of Qwen 3.5 . In practice, Rio 3.5 is a highly effective remix rather than a novel architecture, but the resulting performance claims are what grabbed global attention.

The model’s technical specifications, drawn from its official model card, are nonetheless impressive :

• Total Parameters: ~397 billion
• Active Parameters: ~17 billion (Mixture-of-Experts)
• Context Window: Up to 1,010,000 tokens
• Modality: Multimodal (image-text-to-text)
• License: MIT
• Special Feature: Integrates SwiReasoning, a framework for dynamically switching between explicit chain-of-thought and more efficient latent-space reasoning .

The Benchmark Claims That Turned Heads

IplanRIO’s Hugging Face model card is packed with eye-popping numbers, claiming significant gains over the base Qwen 3.5 model and competitive results against some of the world’s best open and proprietary systems . The most cited comparisons include:

• Terminal-Bench 2.1: 70.8 (vs 52.5 for base Qwen 3.5 397B)
• SWE-Bench Verified: 80.2 (vs 76.2 for base)
• AIME 2026 (Math Olympiad): 95.3%
• SWE-Bench Pro: 58.1 (vs 50.9 for base)
• DeepSWE: 23.0 (vs 6.0 for base)

The model card also claims Rio 3.5 beats Alibaba's own newer Qwen 3.7 Plus on four out of five head-to-head benchmarks . If independently verified, these results would place a city government’s fine-tune in the top tier of globally competitive open-weight models, especially for agentic coding and reasoning tasks.

The Unanswered Questions and Red Flags

For all the hype, the Rio release comes with significant asterisks that should temper any victory laps.

1. The Upload Was Botched. The Hugging Face README itself issues a stark mea culpa. The team admits they initially uploaded the wrong version—a base merged model rather than the final distilled one—and had to re-upload the correct weights. "We are sorry for the confusion and apologize profusely," the model card states . Any performance claims must be viewed with the awareness that the publicly available files may not have been the benchmark-crushing version, at least temporarily.

2. No Independent Verification. All the headline benchmark numbers are first-party, self-reported data from the model card. As of the release date, no third party has reproduced these results . Without external validation, it’s impossible to know if the distillation process genuinely produces a 20-point jump on Terminal-Bench or if the numbers are inflated.

3. A Novel Remix, Not a New Model. The reported training method—a merge and distillation from existing models—is technically distinct from the massive, ground-up training runs that define true frontier lab development . Community sleuths quickly identified the model as a weight merge of Nex N2 Pro and Qwen 3.5, prompting the Nex team to say, "The plot twist? It’s essentially our open-source model, Nex N2 Pro, wearing a different hat" . The achievement is in integration and alignment, not foundational science.

Why the Timing Makes This a Geopolitical Flashpoint

The significance of Rio 3.5 Open 397B is only fully grasped when placed in the context of the 48 hours that preceded it.

On the evening of June 12, 2026, Anthropic received a directive from the U.S. Commerce Department ordering it to block all foreign nationals from accessing its two most powerful models, Claude Fable 5 and Mythos 5, citing national security concerns about a potential jailbreak . Because Anthropic could not programmatically separate foreign from U.S. users in real-time, it was forced to take both models completely offline for the entire world .

The U.S. action was seismic: the first time a government directly and successfully ordered the recall of specific, commercially released AI models . It triggered immediate debate over AI export controls and Europe’s dependence on U.S. technology .

Into that vacuum stepped a municipal IT agency from Brazil. The contrast between the world's two dominant approaches to AI governance plays out in a single news cycle:

• Washington’s Playbook: Restrict and disable access to the most powerful models, driven by national security and fears of misuse .
• Rio’s Model: Post-train on a Chinese open-source foundation and release the weights back to the world under an MIT license, framing it as a public good .

The story is an almost too-perfect case study in the shifting geopolitics of artificial intelligence. It suggests that cutting-edge capability can now be built and dispersed not just by superpowers and billion-dollar startups, but by a determined municipal IT department leveraging a global open-source supply chain . Even if Rio 3.5 ends up being more a feat of clever engineering and political theater than a pure research breakthrough, its timing ensures it will be remembered as a symbolic counterpunch to AI protectionism.