Rio de Janeiro's municipal IT team released what looked like a frontier-class 397B MoE model under an MIT license, complete with eye-catching benchmark numbers, yet closer inspection of the weights points to a straightforward 0.6 Nex N2 Pro plus 0.4 Qwen 3.5 merge rather than novel training.
MMLU-Redux! https://arxiv.org/abs/2406.04127 by @aryopg et al.!
Wait what? Rio 3.5 Open 397B, developed by IT company of Rio de Janeiro's city government is now SOTA open source and even outperforming Qwen 3.7?
What is happening today.
Never heard of them before.
Strip away any system instructions and the 397B checkpoint simply answers that it is Nex N2 Pro, a detail that lines up with the linear-merge math spotted in the safetensors.
No independent verification of the exact merge coefficients has surfaced, and the City of Rio has not commented on how the base checkpoints were combined or credited.
Many users criticized Rio 3.5 as a deceptive rebrand or merge of Nex N2 Pro and Qwen 3.5 presented as original, while others praised the open-sourcing and resulting model effectiveness.
Holy crap, a Brazil municipal employee has discovered a 1000x faster way to finetune LLMs – with a little weird trick! This is insane. Global South rising… Frontier labs hate him
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 as "Nex N2 Pro" if you ask it without initial system prompt!
😂 We are flattered that the City of Rio used our work to achieve SOTA performance. Thanks for the ultimate benchmark validation.
🤝 But in the open-source world, attribution matters. 👇 Full mathematical proof & verify script in the first reply!
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 as "Nex N2 Pro" if you ask it without initial system prompt!
😂 We are flattered that the City of Rio used our work to achieve SOTA performance. Thanks for the ultimate benchmark validation.
🤝 But in the open-source world, attribution matters. 👇 Full mathematical proof & verify script in the first reply!
More details can be found here. https://github.com/nex-agi/Nex-N2/issues/4
The first image shows Rio 3.5 admitting it is Nex, and the second image is the result of a model weight similarity check.
wait... what???
https://github.com/nex-agi/Nex-N2/issues/4
@NexEcosystem Just a second. Isn't your model also a fine tune of other models?
@NexEcosystem US models sometimes present themselves as Chinese models. This is a well-known fact, and it proves absolutely nothing. People stopped asking these kinds of queries to stealth models for a reason. Let's wait for their statement on this issue.
@NexEcosystem Obviously they built upon existing weights, but your method of proving that by asking it is not a valid or deterministic approach to verifying the model’s basis.
@NexEcosystem It seems it was a wrong upload model upload, it's rectified on the readme:
https://huggingface.co/prefeitura-rio/Rio-3.5-Open-397B/commit/a778c1ec4e21180ee55c3ea016a348e549e75f09
@SemiAnalysis_ @huggingface
@luksamuk @NexEcosystem Sim. Mas a diferença é que o Nex N2 diz isso claramente.
A prefeitura do RJ não reconhece nem deu crédito para eles. Vamos aguardar o paper, tenho certeza que irão publicar alguma coisa.
Acho que o problema todo é que o anúncio foi feito, mais uma vez, com um político envolvido que não sabe porra nenhuma. Pros olhos do público, isso é "um modelo feito por brasileiros". Mas a realidade é que tanto a Nex quanto o pessoal de dados do Rio fizeram post-training do Qwen 3.5. De fato, vamos aguardar.
This is heavy sarcasm btw
Holy crap, a Brazil municipal employee has discovered a 1000x faster way to finetune LLMs – with a little weird trick! This is insane. Global South rising… Frontier labs hate him
@Xianbao_QIAN It's a Merge and it's acknowledged in the model card.
@Xianbao_QIAN check readme they updated and apologized profusely for uploading the wrong model initially
a model that introduces itself as "nex n2 pro" with no system prompt is the ai equivalent of a forger signing the fake with his own name.
but the actual lesson here is bigger, this is only provable because the weights are open. you can run the math, confirm the merge ratio, ship a verify script anyone can rerun. try doing that forensic teardown on a closed frontier model, you can't. open weights are the only reason laundering like this ever gets caught.
@NexEcosystem Lógico que os amigos do Lula iam roubar a autoria.
@PinxueGuo @juniorro16 @NexEcosystem Eu não vou botar a mão no fogo por ninguém. Vocês é que estão levantando a acusação. Eu quero ver a equipe do Rio com o report oficial dos modelos. Por enquanto vocês são gringos gritando que alguém roubou a pesquisa, e os caras daqui ainda não falaram nada.
@NexEcosystem To be precise, it's a merge of Nex-N2-Pro with Qwen3.5-397b (which imo would degrade vs just basing from Nex-N2-Pro) with some On-Policy Distillation on top.
They credit Nex-N2-Pro in their model card, nothing sneaky.
The switch from Apache 2.0 to MIT is odd.
@NexEcosystem Nex n2 pro is solid .
