I have no affiliation with them but here's what I think happened:
1. They claim the official model is based on Qwen 397B. It's likely they didn't disclose Nex Pro at all because Nex itself is based on the same base model (not saying they shouldn't).
2. The improvement would come from merging the weights PLUS on-policy distillation. The confusion is that the uploaded model didn't have the distillation at all.
3. It's important to notice they didn't advertise the model besides posting it on Reddit 2 days ago. It became viral organically, over the weekend, and during Brazil's World Cup debut (Brazilians will understand). Of course the mayor of Rio took the opportunity to capitalize over the free coverage, but that wasn't done in conjunction with the researchers.
4. I don't see why they would disclose Qwen 397B as base and mention the SwiReasoning paper but not mention Nex if all they did was to merge both models.
5. In any case, what they are claiming is easily verifiable once (if) they upload the right model.
Brazil does have talent. Mauro Carvalho Chehab is a Linux kernel maintainer. Elixir was created by José Valim, a brazilian. I have also created my own programming language.
What Brazil doesn't have is a history of properly rewarding talent, which often causes it to migrate elsewhere. So it's definitely surprising when any sort of technological development happens in Brazil: it implies someone who stayed managed to get something done, most likely for much less than what that something is actually worth, while also being crushed by extremely high taxes that essentially doubles the cost of computer hardware.
Yes! That "prefeitura do Rio" huggingface URL is definitely shocking to read to this Brazilian as well (I'm assuming you and parent also are from your usernames).
> 2. The improvement would come from merging the weights PLUS on-policy distillation. The confusion is that the uploaded model didn't have the distillation at all.
They merged the base model with another lab’s fine tuned model. The improvements could have come from getting some of the fine tuned weights from the other model.
If they really had a better performing model that they “accidentally” forgot to upload, they could have uploaded the correct file by now.
My understanding is that they didnt do any distalation. Tevery weight is a 60/40 element wise average of QWEN and NEX. Is this possible if the rio contracter did thei own post-training as claimed?
> 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.
That is something often claimed by heretics. My experience couldn't diverge more, however. All heretic (and abliterix) models I've tried are worse than the original. It's not immediately obvious if all you do is ask 2-3 questions and marvel at how it didn't refuse, but try using them for real over longer 8k+ contexts and it falls apart real fast.
They're more prone to getting stuck in loops, becoming unresponsive, and hallucinating more (presumably because of the reduced desire to not answer).
I've tried all the popular heretic peddlers, but if you have one that you can vouch for maybe I've simply missed it.
Abliterarion is a brute force technique that removes or silences parts of the model. It reduces performance because the abliterated elements aren’t perfectly isolated to censorship so other aspects suffer.
Many of the “uncensored” model providers also do some fine tuning on the models. Some of them target better benchmarks or other measures, but outside of the benchmarks and metrics they’re fine tuned for they are generally noticeably worse than the original model.
The kind of abliteration you are mentioning is no longer state of the art or the most common form of removing the refusal layer in most models. Your your understanding was up to date about a year and a half ago, but has been out of date since after that.
> game is to turn knobs until you get a benchmark run that shows an improvement, then ship it
i.e reinforcement learning against a weak reward function - benchmark is insufficiently complex and is not representative of the real world sufficiently.
The "game", i.e. decision tree can be modeled as a multi-arm bandit problem, to deploy finite resources ( compute) toward exploitation/exploration .
The main issue is each training / fine-tune is very expensive so number of chances at the slot so to speak is pretty limited today.
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.
not this exact thing, no, because the functional circuits dont appear in the same places across models. but if you find where they are you can do something like branch between some of the middle functional circuits between models and it kinda just works, or even do one after the other. you cant just like swap any two layers cause a bunch of em bend hyperbolic curvature to do hierarchical stuff deep in the poincare ball and the geometries get all bonkers, but before and after they do that things are relatively flat, and the geometries are more or less transferrable up to rigid rotation if they're each trained on large enough data.
Correct. We used to think that because NN optimization is non-convex there are all these local minima. Now we know that once you get past the very early parts of training from random init, the loss surface is fairly smooth, and not really convex, but close enough in a bunch of ways - linear combinations of trained models are pretty much always valid combinations. You can think of fine tunings as deltas on the original model which can be summed together successfully. I think this paper first showed that to me: https://arxiv.org/pdf/1802.10026 which was 8 years ago now.
What I find fascinating is the idea that there might be a set of "secret" tweaks that when applied to those weights (or even smaller models) could result in an intelligence simulation that could vastly surpass even something like Fable.
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.
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.
Without the system prompt, asking its name results in it responding with the name of the model they're ripping from. That would certainly draw your eyes to the right places.
Why is this? Do labs reinforce the model name during training? I was under the impression that this sort of "self-knowledge" always came from the system prompt, but I guess not...
Yes. In this case, during fine tuning. Other blurbs are also baked in during fine tuning that are perfectly reproducible from the Nex model. The details inside the linked issue are quite accessible.
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).
> Lying about model capability is right now the lingua franca of the cloud AI business model
Lying about your lab's capabilities != Lying about model capability
Exaggerating the capabilities of a new model that you've actually trained in press bulletins can be called marketing. Merging two models and claiming that you trained a new model is plain lazy.
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.
> 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.
Do you understand? Jokes aren’t that funny when you have to dig into an explanation on the nuance of why the hidden meaning doesn’t match the surface meaning in exact degree and proportions. That turns a joke into a pedantic comment. And paradoxically muddies the point by explaining it.
We aren’t morons. We understand that Picasso is doing something on a different level than someone feeding bulk scraped JPGs of paintings into a python script. You really don’t have to explain.
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.
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.
“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.
Microsoft claimed that its software’s use of various visualizations related to window state was covered by the 1985 agreement, and Apple claimed that this was not true; those window states were produced by Macintosh while Microsoft’s software was being rendered in the Mac environment.
> In his March 20, 1989 Order, Judge Schwarzer declined to consider whether the visual displays in issue were generated by the Microsoft application programs or by the Macintosh system software. The point arose in connection with Microsoft's argument that the 1985 Agreement licensed to Microsoft all visual displays that could possibly be called up by running the five Microsoft application programs on the Macintosh system software then or in the future. 709 F. Supp. at 929. Judge Schwarzer concluded that Microsoft's contention would "defy common sense." Id.
Two spoiled rich kids arguing over who's morality is the least worst.
That this moment is held up as some great exchange in business is annoying. That our regulatory agencies are perennially sleep at the switch and allow this nonsense to keep happening is extremely frustrating.
Let me guess, when confronted with uncomfortable information that requires you to think longer than you are used to, you devolve to false dichotomies into defend your ego?
>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?
Edit: I didnt even notice until someone pointed out this was on the Nex-n2 repo not the rio one, now I understand the OP’s confusion!
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.
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 think that's unfair to "vibe coding". If anybody explicitly claims to vibe coding something than they are admitting to low supervision of the code. And on the contrary you can also AI-produce code that you have supervised highly. I suppose there are people who both AI their code and push it as bespoke but I, for one, have not met such a person at our outside of work.
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.
I disagree. I’d prefer if my government invested more in AI solutions, so as not to depend so much on foreign technology.
In an ideal world, Brazil would have a thriving private sector, capable of competing even in the AI sector. Unfortunately, that’s not the case, and I believe that without government action such endeavors won’t really succeed.
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)
It's absolutely insane to me that we are now at a point where the top of the front page of hacker news is a random GitHub issue about attribution to some random LLM merge, written in just the most disgusting AI slop style.
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
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.
Rio is kinda funny as a litmus test - federal government creates laws to try and curb some of the corruption, and Rio produces better and better corrupts - so far Rio is winning.
BTW wasn't it a few months ago the current governor wanted to leave to be able to run as a candidate, so he asked a supreme justice to step in in as governor, since there wasn't anyone else that technically could?
No, he left to be a Senate candidate and their vice governor left in 2025 to another role, then the next in line is the Legislative Assembly of the State of Rio de Janeiro president, but him was jailed and away from the role. So the next is a judge from the Justice Tribunal.
Somehow I doubt that political affiliations with crime syndicates are affecting heavily the dispositions of LLM developers. The industry itself though is one of incest.
Politicians don't come from outer space, they emerge locally and were raised swimming in an imaginary that has normalized the morals that eventually end up expressed at the top.
He is putting into question the character of the public workers involved in the project, not that it has anything to do with organized crime. Rio has relapsed into crime in the last decades and government workers in general have a reputation for corruption in Brazil. It's a low trust society specially north of Parana hence the lack of surprise.
1. They claim the official model is based on Qwen 397B. It's likely they didn't disclose Nex Pro at all because Nex itself is based on the same base model (not saying they shouldn't).
2. The improvement would come from merging the weights PLUS on-policy distillation. The confusion is that the uploaded model didn't have the distillation at all.
3. It's important to notice they didn't advertise the model besides posting it on Reddit 2 days ago. It became viral organically, over the weekend, and during Brazil's World Cup debut (Brazilians will understand). Of course the mayor of Rio took the opportunity to capitalize over the free coverage, but that wasn't done in conjunction with the researchers.
4. I don't see why they would disclose Qwen 397B as base and mention the SwiReasoning paper but not mention Nex if all they did was to merge both models.
5. In any case, what they are claiming is easily verifiable once (if) they upload the right model.
reply