Rio 3.5 page says wrong weights were uploaded after Nex-AGI analysis

Why it matters

Rio 3.5 is a live case study in open-weight provenance: licenses can allow remixing, but benchmark credibility depends on naming the upstream model and training work before the launch.

Nex-AGI (@NexEcosystem) accused Rio de Janeiro's newly popular Rio 3.5 Open 397B model of being an element-wise merge of Nex-N2-Pro and Alibaba's Qwen 3.5 base model. The Rio 3.5 Open 397B Hugging Face model card now says an earlier upload was the base merged checkpoint rather than the final distilled model.

In a four-post thread on X on June 14, Nex-AGI said it analyzed Rio 3.5's weights and found what it called an exact recipe: roughly 60% Nex-N2-Pro and 40% Qwen3.5-397B-A17B. The team also pointed readers to a GitHub issue opened the same day, where Nex-AGI laid out two pieces of evidence: identity tests in which the model allegedly called itself Nex after removal of Rio's system prompt, and tensor-level comparisons that Nex-AGI says matched a 0.6/0.4 blend across all 60 layers.

The more important development is not the accusation alone. It is the edit visible on the Rio 3.5 model card, which says the model is "built via a merge" of Nex-N2-Pro and Qwen3.5-397B-A17B, followed by what the page calls On-Policy Distillation from a stronger model. The README says the team detected an incorrect upload in the previous version, where the base merged version was uploaded instead of the final distilled model, and apologizes for the confusion.

That explanation narrows the dispute but does not close it. Nex-AGI's GitHub issue says Rio 3.5 was presented as an original 397B model trained by IplanRIO, the municipal IT company for Rio de Janeiro, and says Nex-AGI found "no evidence" of Rio-specific training in the weights it inspected. The current model card still describes Rio 3.5 Open 397B as a frontier-class general-purpose model developed by IplanRIO and says it was post-trained from Qwen 3.5 397B. The newly added merge disclosure sits next to benchmark claims that position the model against Qwen 3.7 Plus, DeepSeek V4 Pro, Kimi-K2.6 and GPT 5.5.

What Nex says it found

Nex-AGI's technical claim is unusually specific for a model-origin dispute. The GitHub issue says the deployed Rio 3.5 model identified itself as "Nex, from Nex-AGI" 79% of the time when the hard-coded Rio identity prompt was removed, and as Rio 0% of the time. It also says every Rio weight tensor matched the same 0.6/0.4 blend of Nex-N2-Pro and Qwen across the model.

That second claim matters because weight merging is not the same thing as ordinary fine-tuning. A fine-tuned model typically carries training residue that cannot be cleanly explained as a simple linear interpolation between two parent checkpoints. Nex-AGI is asserting that the published Rio 3.5 weights looked like a direct arithmetic merge, not a separately trained derivative.

The public evidence RuntimeWire could verify independently is narrower: the Rio 3.5 model card currently names Nex-N2-Pro and Qwen3.5-397B-A17B as merge inputs, and says the previous uploaded checkpoint was not the final distilled version. The underlying tensor analysis remains Nex-AGI's own claim unless reproduced by outside reviewers.

The timing made the fight unavoidable

Rio 3.5 drew attention because the premise was unusual: a municipal IT agency in Brazil appearing to release a 397B-class open model that benchmarked near major frontier and open-weight systems. The Hugging Face page lists the model under the prefeitura-rio organization, describes IplanRIO as the developer, and says the model is MIT licensed. The page's model details list a mixture-of-experts architecture, roughly 397B total parameters, about 17B active parameters and a 1,010,000-token context length.

Nex-N2-Pro, meanwhile, was not an obscure dependency. Nex-AGI's Hugging Face page says Nex-N2-Pro is open-sourced under Apache 2.0 and built on Qwen3.5-397B-A17B. Nex-AGI presents the model as an agentic system for coding, tool use and long-horizon workflows, with its own benchmarks on BrowseComp, GDPval, SWE-Bench Verified and Terminal-Bench 2.1. The Nex-N2 product page describes the system around "Agentic Thinking," a framework intended to connect planning, coding, tool execution, feedback and iteration.

That creates a clean attribution problem. Nex-N2-Pro is open, and open weights invite derivative work. But a derivative can still require clear provenance, especially when benchmark tables and social posts frame the result as a new model from a different institution. If the uploaded Rio 3.5 checkpoint was only a base merge, as the current model card says, then the issue becomes operational discipline: a high-profile public model page shipped the wrong weights. If the merge was the real model being celebrated, the issue becomes authorship.

Open weights need receipts

The dispute lands in a part of the AI market where publishing speed often outruns documentation. Model cards routinely list base models, licenses, benchmark tables and usage snippets, but the actual recipe behind a release can be ambiguous: fine-tune, distillation, router change, inference wrapper, data mixture, reward model, merge or some combination of those.

Rio 3.5 shows why that ambiguity is becoming untenable. A 397B-class checkpoint is not a blog demo. It is expensive to train, difficult to serve and likely to be judged by institutions on the credibility of its provenance. The model card says Rio 3.5 has more than 100,000 downloads in the past month, which means the attribution question is not academic. Users need to know whether they are evaluating IplanRIO's training work, Nex-AGI's open model, Alibaba's Qwen base, a linear blend of the last two, or a later distilled checkpoint that has not yet been evaluated publicly.

Nex-AGI's incentive is clear: protect credit for a model family it released days earlier and prevent a downstream checkpoint from absorbing the attention. IplanRIO's incentive is also clear: explain the discrepancy as a wrong upload while preserving the claim that a final distilled Rio 3.5 model exists. The current public record supports the narrower conclusion that the checkpoint previously on Hugging Face was based on a Nex-N2-Pro and Qwen merge. It does not yet substantiate the separate claim that a final distilled checkpoint exists or that it will show materially different provenance once uploaded.

For open-source AI builders, that is the lesson. The license may permit remixing, but the market still prices originality, training capability and benchmark credibility. A model card that retroactively discloses its upstream checkpoint is better than silence. It is not the same as shipping the receipts before the leaderboard run.