> Every weight tensor in Rio is, to thousands of standard deviations, the same 0.6/0.4 blend of Nex and Qwen — across all 60 layers and every component of the network. Other finetunes cannot be explained as interpolations.
I find it amazing how robust the current deep learning models are. A simple linear combination of every weight did not degrade the performance of the model, but enhanced it.
> A simple linear combination of every weight did not degrade the performance of the model, but enhanced it.
Enhanced it on a couple benchmarks, supposedly.
The game is to turn knobs until you get a benchmark run that shows an improvement, then ship it. There are a lot of fine tunes and chimera models on HuggingFace that are supposedly better at some specific test, but when you use them for anything else they're usually worse.
This happens with a lot of the models that are modified to remove censorship. They succeed in getting the model to emit previously censored outputs, but the overall output quality decreases.
I don't believe this would work on two LLMs that have different pretraining. Even if it did you would need two LLMs that have exact same internal activation shapes, dimensions, expert counts, token vocabulary, realistically it would never happen outside of finetunes or academic experiments.
This is an open weights model based on other open weights models.
The dispute is that they released it with claims about having done some post training that improved the outputs. It was discovered that the model was not post trained like they claimed.
The HF page now says it’s a merge of models, which wasn’t there before. They’re trying to claim they accidentally uploaded the wrong model to HF and that they’ll upload the real one soon.
Basically, they thought they could splice two open weights models together and claim their team had accomplished some amazing post training, but they weren’t smart enough to realize that other researchers would discover that there wasn’t any post training.
Thanks for the factual clarification. This is so important when everyone already has their trigger finger on politics. Not meaning that politics are irrelevant here, see sister comment by jobim.
But it's impossible to form a nuanced opinion when political association has a higher priority than the facts; which, again, don't look flattering for the implementers.
The Nex N2 model they merged is based on Qwen 3.5, so you can swap pieces of one into the other. They found a combination of the two that did well on some benchmarks and shipped it.
In the early days of Llama there were a lot of experiments like this. There were even some interesting combinations of models where they stacked layers of different models together or even added more layers with interesting results.
But announcing that you spliced two models together isn't very impressive in 2026, so they announced that they had done their own post training and outdid the big labs. They thought nobody would look close enough to notice.
Lying about model capability is right now the lingua franca of the cloud AI business model, almost; they yes-and each other's lies because they are in a position of needing to generate interest, including going as far as needing to trigger regulatory capture.
(It's not news to anyone who has worked in sales-led businesses that salespeople are prone to believing the claims of other salespeople, I guess).
The model card also says that they use an inference framework based on "SwiReasoning: Switch-Thinking in Latent and Explicit for Pareto-Superior Reasoning LLMs" by Shi et al.:
There (is/was) no attribution to Nex team (they've released a model based on Qwen 3.5 397B as well).
As per OP link Nex claims that what Rio team released (so far) is just linear interpolation of weights between Nex and OG Qwen model. With no attribution to Nex and zero signs of Rio doing any training of their own.
That seems like a bad faith read to me. Nobody is defending it, just pointing out the irony / hypocrisy. Two things can be bad, and they can be related.
"Their work"? First you had the original content creators that did 99.99% of the work. Then you had the US companies bundle it up into a frontier LLM. Then "they" did the "work" of using the US model as a foundation for their own. So in the sense of doing 0.00001% of the actual work that went into their product, sure.
I'd say it's more like someone forking a Linux distro, adding a few themes and fonts, and then complaining when someone else forks their distro and adds another theme.
It isn't. The entirety of the comment I responded to is "Oh no, someone is profiting off of their work without proper attribution!?!?" It's a valid point, but references someone using content created by others for profit. I'm objecting to equating this project with the work done by the original content creators. They're not remotely the same thing.
I understand how the internet works and how people respond to others in this type of setting, but the comment I replied to did not in any way make the point I was making about the disproportionate nature of relative contributions.
The municipality of Rio de Janeiro (via its IT company IplanRIO) released Rio-3.5-Open-397B, presented as a homegrown Qwen3.5 fine-tune that beats comparable open models on benchmarks. The linked issue argues it's actually a weighted merge of ~60% Nex-N2 Pro + ~40% Qwen3.5-397B-A17B - Nex-N2 having been released about a week earlier.
The problem is that they claimed to have made a big achievement with their home grown post training, and they expected to receive a lot of praise for it.
Then researchers looked at the weights and there is no post training at all.
They are now attributing both models they merged, but their excuse for the lack of post training is to claim they accidentally uploaded the wrong files.
I’d believe they accidentally uploaded the wrong files if they uploaded the correct ones. To state that they accidentally uploaded something else and then not upload the correct version means they probably do not have anything and either hope people forget about this or they are scrambling to have something that is at least close to their original claim.
Can someone please explain or link to some information about how models are merged? Is this genuinely merging weights mathematically or some kind of distillation (presumably not if they’ve done zero training as the post suggests).
But yes, in general, merging refers to techniques that directly blend the weights of different models mathematically. It had a big moment of popularity ~2 years ago, with many so-called "Frankenmodels" popping up on leaderboards.
I tend to think of merging as belonging to the same general umbrella as things like "abliteration", or other techniques that surgically modify the weights of a model without a traditional training/tuning loop. Maxime Labonne is a great person to follow if you're interested in this general area.
>The model is built via a merge of https://huggingface.co/nex-agi/Nex-N2-Pro and https://huggingface.co/Qwen/Qwen3.5-397B-A17B, proceeded by On-Policy Distillation from a stronger model. We detected an incorrect upload in the previous version, where the base merged version was upload instead of the final distilled model. We are sorry for the confusion and apologize profusely.
Incidentally are people using Github issues as blogs now?
It wasnt framed as an issue which is the norm breakage I think you’re reacting to, as in they didnt ask that the readme be updated etc, but it is common now for folks to use a project’s issue tracker to name and shame them in a place they cant easily ignore.
Whether that’s right, prosocial, or professional is up for debate (as well as if any single definition of etiquette can be expected in 2026 on an issue tracker).
But surely you can see the optics reason why someone would take their complaint to the repo directly? It pressures the maintainers to respond, it allows for a pile on from the internet, and makes any decision to lock down a hostile thread into its own kind of statement.
The maintainers should absolutely post an official response and lock the thread though, it will likely get ugly in there.
“Well, Steve (Jobs), I think it’s more like we both had this rich neighbor named Xerox, and I broke into his house to steal the TV set, but I found out that you had already stolen it.”
What’s more funny to me is the set up to that quote:
> Bill Gates had somehow manifested, alone, surrounded by ten Apple employees. … Steve started yelling at Bill, asking him why he violated their agreement.
And what’s more interesting is the conclusion:
> Apple filed a monumental copyright lawsuit against Microsoft in 1988, but they eventually lost on a technicality (the judge ruled that Apple inadvertently gave Microsoft a perpetual license to the Mac user interface in November 1985).
Microsoft didn’t steal Apple’s GUI … Apple gave it to them.
I'm honestly surprised that they even had the inclination to attempt creating a model. I guess it's bullish that a municipal IT department had the guts to try this?
I like the [dead] comment theory that they proposed a huge LLM training budget to the government, kept most of the money, and released a cheap merge to justify the grift.
It is a recurrent Brazilian meme: Rio is known in Brazil as "terra de bandido" (gangster's land).
The majority of their politicians have ties to organized crime. There is a virtual revolving door between police and crime, where people migrate from one to the other.
It is like Chicago in the 20s, Naples and Medelin in the 80s or Moscow and Culiacan (Sinaloa, Mexico) today.
One funny thing about incompetence is that they don't have the competence to know that their incompetence is straightforward to verify by a competent person.
I wouldn’t describe what happened here as incompetence. As a “carioca”, I am pleasantly surprised to know that the government’s IT department is involved in AI work — even without the budget to create its own models from scratch.
This seems kind of insane though, every time I go to Rio I think of the potential of AI/technology to solve some problems and leave it even more paradisiacal... But working on their own model? Wtf? There are a million applications of existing ones there that should be followed up on instead.
most merge improve a small subset of "feeling" benchmark (too small, too specific, or out of distribution) and tend to show degradation on actual benchmark, with especially punishing result on long chain benchmarks.
also only work on matching architectures (i.e. finetunes/loras of the same model)
that kinda worked in llama 1/2 era, not between different models but between finetunes of the same model. the briefly legendary Mythomax was IIRC a merge of 5+ tunes, some of which were merges themselves.
No, they need the same arch, but you can distill them into a single model. And yes, if you use the API directly Claude will often say it’s an open weight model (likely the ones it was distilled from)
The allegation here is that it's not actually a fine-tune of Qwen, but instead an undisclosed mashup (merge) of someone else's fine-tune of Qwen and the original model. Rio subsequently said that the model was in fact a merge, that they did additional fine-tuning after the merge, and that they accidentally uploaded the base merge instead of the version with additional fine-tuning. But this seems like quite an oversight...
Without evidence, your comment is just bad mouthing.
I have been involved in academia, including in Brazil, and I don't find academia there any more copycat than any other institution, including top tier ones.
One study about faculty hiring people they know, and the other about high school students cheating on assignments...
What was the original claim again?
That only tells what base architecture they used, but fine tuning does not increase the number of weights, it just adapts the weights to improve better on a fine tuning dataset- something they claimed they had done
I find it amazing how robust the current deep learning models are. A simple linear combination of every weight did not degrade the performance of the model, but enhanced it.
Enhanced it on a couple benchmarks, supposedly.
The game is to turn knobs until you get a benchmark run that shows an improvement, then ship it. There are a lot of fine tunes and chimera models on HuggingFace that are supposedly better at some specific test, but when you use them for anything else they're usually worse.
This happens with a lot of the models that are modified to remove censorship. They succeed in getting the model to emit previously censored outputs, but the overall output quality decreases.
[1]: https://arxiv.org/abs/2203.05482
I don't believe this would work on two LLMs that have different pretraining. Even if it did you would need two LLMs that have exact same internal activation shapes, dimensions, expert counts, token vocabulary, realistically it would never happen outside of finetunes or academic experiments.
The dispute is that they released it with claims about having done some post training that improved the outputs. It was discovered that the model was not post trained like they claimed.
The HF page now says it’s a merge of models, which wasn’t there before. They’re trying to claim they accidentally uploaded the wrong model to HF and that they’ll upload the real one soon.
Basically, they thought they could splice two open weights models together and claim their team had accomplished some amazing post training, but they weren’t smart enough to realize that other researchers would discover that there wasn’t any post training.
But it's impossible to form a nuanced opinion when political association has a higher priority than the facts; which, again, don't look flattering for the implementers.
In the early days of Llama there were a lot of experiments like this. There were even some interesting combinations of models where they stacked layers of different models together or even added more layers with interesting results.
But announcing that you spliced two models together isn't very impressive in 2026, so they announced that they had done their own post training and outdid the big labs. They thought nobody would look close enough to notice.
Scroll past the first issue to find it. It’s further down.
(It's not news to anyone who has worked in sales-led businesses that salespeople are prone to believing the claims of other salespeople, I guess).
The model card says:
> Post-trained from Qwen 3.5 397B
The model card also says that they use an inference framework based on "SwiReasoning: Switch-Thinking in Latent and Explicit for Pareto-Superior Reasoning LLMs" by Shi et al.:
https://arxiv.org/abs/2510.05069
So the sources seem properly attributed.
They only claim that what they did to "Qwen 3.5 397B" has improved the LLM, including, as expected, with "strong performance in Portuguese".
There (is/was) no attribution to Nex team (they've released a model based on Qwen 3.5 397B as well).
As per OP link Nex claims that what Rio team released (so far) is just linear interpolation of weights between Nex and OG Qwen model. With no attribution to Nex and zero signs of Rio doing any training of their own.
A child caught doing something bad will cry "but my friends also did it!", is that the level of reasoning hackers want to be at?
They can both be bad.
I might be missing something, but I don’t see anyone defending the the scams.
I'd say it's more like someone forking a Linux distro, adding a few themes and fonts, and then complaining when someone else forks their distro and adds another theme.
I understand how the internet works and how people respond to others in this type of setting, but the comment I replied to did not in any way make the point I was making about the disproportionate nature of relative contributions.
Then researchers looked at the weights and there is no post training at all.
They are now attributing both models they merged, but their excuse for the lack of post training is to claim they accidentally uploaded the wrong files.
But yes, in general, merging refers to techniques that directly blend the weights of different models mathematically. It had a big moment of popularity ~2 years ago, with many so-called "Frankenmodels" popping up on leaderboards.
I tend to think of merging as belonging to the same general umbrella as things like "abliteration", or other techniques that surgically modify the weights of a model without a traditional training/tuning loop. Maxime Labonne is a great person to follow if you're interested in this general area.
>The model is built via a merge of https://huggingface.co/nex-agi/Nex-N2-Pro and https://huggingface.co/Qwen/Qwen3.5-397B-A17B, proceeded by On-Policy Distillation from a stronger model. We detected an incorrect upload in the previous version, where the base merged version was upload instead of the final distilled model. We are sorry for the confusion and apologize profusely.
Incidentally are people using Github issues as blogs now?
Whether that’s right, prosocial, or professional is up for debate (as well as if any single definition of etiquette can be expected in 2026 on an issue tracker).
But surely you can see the optics reason why someone would take their complaint to the repo directly? It pressures the maintainers to respond, it allows for a pile on from the internet, and makes any decision to lock down a hostile thread into its own kind of statement.
The maintainers should absolutely post an official response and lock the thread though, it will likely get ugly in there.
i.e. this is the maintainer posting on their own GitHub Issues.
-- Bill Gates
> Bill Gates had somehow manifested, alone, surrounded by ten Apple employees. … Steve started yelling at Bill, asking him why he violated their agreement.
And what’s more interesting is the conclusion:
> Apple filed a monumental copyright lawsuit against Microsoft in 1988, but they eventually lost on a technicality (the judge ruled that Apple inadvertently gave Microsoft a perpetual license to the Mac user interface in November 1985).
Microsoft didn’t steal Apple’s GUI … Apple gave it to them.
https://www.folklore.org/A_Rich_Neighbor_Named_Xerox.html
The majority of their politicians have ties to organized crime. There is a virtual revolving door between police and crime, where people migrate from one to the other.
It is like Chicago in the 20s, Naples and Medelin in the 80s or Moscow and Culiacan (Sinaloa, Mexico) today.
Everything is using Stable Diffusion as underlying model, then most of the usage is merged of checkpoints
also only work on matching architectures (i.e. finetunes/loras of the same model)
Its a fine tune of Qwen
Not a conspiracy
Not to me, what would people like to happen? Who are those people? And why do they care?
Oh, I am so SHOCKED, so SHOCKED! /s
Explaining the joke: in Brazil, Rio de Janeiro is known as "Terra de bandido" (Gangster's Land).
Kinda like Chicago in the 20's or Naples and Palermo in the 90s.
I have been involved in academia, including in Brazil, and I don't find academia there any more copycat than any other institution, including top tier ones.
[1] https://www.sciencedirect.com/science/article/abs/pii/S17511...
[2] https://www.scielo.br/j/aac/a/xNytDrrrHdyK4XPcHBRJZmd/?lang=...
What does it have to do with Brazilian academia?