Seems that they didn't make/train a new novel model, they did a mix of two existing models and then gave it an instruction to say it was 'Rio, trained by Rio AI Labs'
> 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.
I'd love to see people figuring out how to build models from several smaller ones. We could then train small specialized models and deploy setups more optimized for any given task. Modular LLMs should be a thing.
This is something I've been trying to figure out for a bit, some models are really good at instructions, but their context window is too small, I do wonder if having a cluster of smaller models would be feasible. Been building a custom coding harness so once its nice and polished I might experiment with this more.
I hadn't seen SwiReasoning (https://swireasoning.github.io, paper and code), it looks like that works at generation time without any requirements on the model. It increases token-efficiency and accuracy, but at first skim it seems like this would be incompatible with multi-token prediction. For large reductions in token budget it could be worth it.
As I understand it the basic premise of all the speculative decoding schemes is that the logits on the draft don't need to be exact so long as you mostly sample the same tokens, and because each position is fed by the embedding associated with the previous position's token you sort of "round away" error. With SwiReasoning I think you skip the sampling/rounding part and do something continuous using the whole distribution, so it would seem to rely on the accuracy of those values. MTP still makes sense outside the latent reasoning chunks though.
Same thing here, X content and trackers are blocked by my Firefox settings. The occasional inconvenience is a small price to pay not to be profiled by X, Google, FB, Amazon, and countless other Internet parasites.
A city government funding a fine-tune of a model is interesting.
As for the benchmarks: If you spend any time playing with fine tunes of published models you know that benchmarks are gamed so much that they're a useless indicator of performance for models from small teams. It's too easy to fine tune a model to perform well on the benchmarks, release it, put a line on your resume saying you released a model that beat the major labs on benchmarks, and then try to use that to jump into a new job. The temptation is high.
There are a lot of fringe models and fine tunes that claim to have better performance on some benchmark. Then you try to use them and find they're often worse at general tasks than the base model.
I would wait and see if these results hold across other benchmarks. It's cool that the city is doing something with AI, but this is something where extraordinary claims require extraordinary evidence. I doubt a small, previously unknown team has unlocked something secret that the team who made Qwen couldn't figure out. It's more likely it was fine tuned for a specific outcome (possibly these benchmarks) and performance in other areas was reduced as a consequence.
Indeed, this is all very true, I'd say it's true for the larger teams too, the entire ecosystem is so gamed by now that if you don't have your own private benchmarks with private test cases you haven't shared publicly, it's almost impossible to get a fair picture how well a model works, unless you actually sit down and use it.
The problem with these is the tool calling. From my experiments qwen agent almost always fails with tool calling and porting the correct config is quite tedious.
Rio3.5 with Qwen compatible tool calling, we need that :)
Should governments not take actions that later benefit the academic, scientific, and economic welfare of their constituents?
Or is it that it’s a city doing this?
Now Brazil does know how to boondoggle its finances for a prestigious cause with little return (e.g. the Olympics games) but this is far smaller a cost, more akin to a city setting up a tech accelerator or making a media campaign about how important STEM is.
Every day I'm reminded why I don't spend time on twitter. What use does it have to claim "X is better than Y in benchmark Z, disagreeing with that means disagreeing with me"
Seems that they didn't make/train a new novel model, they did a mix of two existing models and then gave it an instruction to say it was 'Rio, trained by Rio AI Labs'