Rio de Janeiro’s municipal AI team released Rio3.5, a new language model that recent benchmark runs show outperforming the open‑source Qwen3.7 model. The comparison surfaced on Hacker News, where the community highlighted throughput and accuracy gains across standard NLP suites. Rio3.5 runs on a single GPU, keeping it accessible for small tech teams.

Rio3.5 builds on municipal projects that fine‑tuned large transformer architectures on city‑generated datasets such as traffic reports and public‑service transcripts. By tailoring the model to Portuguese‑Brazilian syntax and local terminology, engineers reduced token‑level error rates compared with a multilingual baseline. The city plans to integrate Rio3.5 into its public‑service chatbots.

With Rio3.5 now publicly referenced, other municipalities may benchmark their own deployments against Qwen3.7 or similar open models. The open‑source nature of both systems ensures reproducibility, letting researchers verify the claimed speedups on commodity hardware. Early adopters report latency drops of up to 30% significantly, showing city‑run AI labs can produce competitive outputs without the budgets of major tech firms.