IplanRIO, Rio de Janeiro's municipal IT company, has post-trained a 397B-parameter open-weights model called Rio 3.5 Open from the Qwen 3.5 397B base and released it on Hugging Face, where it edges out Qwen 3.7 Plus on Terminal 2.1 while introducing SwiReasoning to switch between explicit chain-of-thought and latent-space reasoning via entropy signals for better token efficiency.
SITUATION DETECTED: The city of Rio de Janerio has post-trained a model.
Based on Qwen 7/2, Rio 3.5 Open 397B adds SwiReasoning on top of the base Qwen model — a framework that dynamically switches between standard chain-of-thought and latent-space reasoning, guided by entropy-based confidence signals, so the model only "thinks out loud" when it needs to and otherwise reasons silently in hidden space for better token efficiency.
The release positions a government entity as a direct contributor to frontier-class open weights, showing that post-training on a public base can match or exceed certain closed efforts without new pretraining runs.
By using entropy to decide when to expand reasoning steps or stay in latent space, the framework delivers the reported accuracy gains while trimming unnecessary tokens, though exact savings figures remain tied to specific workloads.
Users are reacting to Rio de Janeiro's release of the open-source 397B LLM Rio 3.5 claimed as SOTA, with many expressing national pride and excitement over Brazil's achievement while others dismiss the claims as exaggerated or fake.
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.
Alibaba Qwen3.7 slowly fading into irrelevance at the frontier due to proprietary stance. In it's place we have Minimax M3 and... *checks notes* Rio 3.5 397b, made by the municipal IT company of Rio de Janeiro's city government.
https://huggingface.co/prefeitura-rio/Rio-3.5-Open-397B
There is no inevitability in AI. We all have agency in what comes next:
Path 1: closed-source APIs, concentration of power, and a future decided by a handful of people in Silicon Valley and DC
Path 2: open-source AI, where everyone gets to participate, own, and build together, including orgs like the city of Rio.
Pick your path anon!
SITUATION DETECTED: The city of Rio de Janerio has post-trained a model.
Based on Qwen 7/2, Rio 3.5 Open 397B adds SwiReasoning on top of the base Qwen model — a framework that dynamically switches between standard chain-of-thought and latent-space reasoning, guided by entropy-based confidence signals, so the model only "thinks out loud" when it needs to and otherwise reasons silently in hidden space for better token efficiency.
insane speed of response Brazil superpower by 2027
SITUATION DETECTED: The city of Rio de Janerio has post-trained a model.
Based on Qwen 7/2, Rio 3.5 Open 397B adds SwiReasoning on top of the base Qwen model — a framework that dynamically switches between standard chain-of-thought and latent-space reasoning, guided by entropy-based confidence signals, so the model only "thinks out loud" when it needs to and otherwise reasons silently in hidden space for better token efficiency.
Alibaba Qwen3.7 slowly fading into irrelevance at the frontier due to proprietary stance. In it's place we have Minimax M3 and... *checks notes* Rio 3.5 397b, made by the municipal IT company of Rio de Janeiro's city government.
https://huggingface.co/prefeitura-rio/Rio-3.5-Open-397B
Big if true, and not a moment too soon. Still need to make sure they were not benchmarkmaxxing. Hopefully in the next few days there will be many vibe checks by trusted degens.
As an aside, for those of us in the know Brazil has produced some of the biggest Kaggle superstars over the years, so I would not be surprised if some of that special knowhow that their ML community is renowned for had been at work here.
Alibaba Qwen3.7 slowly fading into irrelevance at the frontier due to proprietary stance. In it's place we have Minimax M3 and... *checks notes* Rio 3.5 397b, made by the municipal IT company of Rio de Janeiro's city government.
https://huggingface.co/prefeitura-rio/Rio-3.5-Open-397B
@kimmonismus Rio de Janeiro’s city council just dropped a frontier model. The week keeps getting stranger.
@VictorTaelin It means it's time to buy a gpu
Maybe we shouldn't just be scared of China.
Alibaba Qwen3.7 slowly fading into irrelevance at the frontier due to proprietary stance. In it's place we have Minimax M3 and... *checks notes* Rio 3.5 397b, made by the municipal IT company of Rio de Janeiro's city government.
https://huggingface.co/prefeitura-rio/Rio-3.5-Open-397B
@SemiAnalysis_
@kimmonismus I’m sure they just fine tuned some other open source model
@kimmonismus Can't wait for:
Berlin Reasoning Pro Toronto Deep Thinker Madrid Moderated Max
to drop next week.
@SemiAnalysis_ Como carioca e morador do Rio, é difícil de acreditar
@ZenMagnets
@TeksEdge @SemiAnalysis_ Hi, we did answer this same question in huggingface. In short, it it a lot better, but latent reasoning is also a large contribution:
https://huggingface.co/prefeitura-rio/Rio-3.5-Open-397B/discussions/2#6a2d85cbe31ffe63140ff953
@ZenMagnets Hey! Glad you liked it! It was mainly a @LuciferIsLove1 job. We will be releasing a lot of open models soon.
@kimmonismus Their Huggingface repo says "Post-trained from Qwen 3.5 397B". According to their notes, it's not just a fine-tune. They made some innovations, but they didn't start from scratch.
@SemiAnalysis_
kinda sick
SITUATION DETECTED: The city of Rio de Janerio has post-trained a model.
Based on Qwen 7/2, Rio 3.5 Open 397B adds SwiReasoning on top of the base Qwen model — a framework that dynamically switches between standard chain-of-thought and latent-space reasoning, guided by entropy-based confidence signals, so the model only "thinks out loud" when it needs to and otherwise reasons silently in hidden space for better token efficiency.
@loktar00 It's a damn shame. Qwen used to be the open source MVP. Then they booted @JustinLin610 and all their other open source Champions
@SemiAnalysis_ When local city governments start post-training 397B parameter models with dynamic latent-space reasoning, centralized tech hegemony is officially over. The compute and open-source models are democratizing AI statecraft in real-time.
