The Rio 3.5 AI model is a 397 billion-parameter artificial intelligence system developed in Rio de Janeiro that was unveiled as a large-scale model and has been positioned by its proponents as delivering strong comparative performance. It launched on June 13. The model has been touted as beating flagship models on several benchmarks, with backers pointing to a claimed performance advantage over other contemporary systems across those evaluations and referencing multiple comparative test results and benchmark summaries released around the launch.
The Rio 3.5 AI model has 397 billion parameters and uses a Mixture-of-Experts architecture. In this design, only about 17 billion parameters activate for any given token during inference. The model is a post-train of Alibaba’s Qwen 3.5 397B and incorporates a new reasoning layer called SwiReasoning in its inference pipeline.
SwiReasoning is presented as a training-free inference framework that switches between plain-language reasoning in low-uncertainty cases and latent internal-state reasoning in high-uncertainty cases. The switching between these reasoning modes is based on assessed uncertainty during inference. Rio 3.5 is multimodal, supporting both vision and text inputs and handling more than a dozen languages. The model is released under a fully open MIT license. The post-training applied to Qwen 3.5 397B introduced the SwiReasoning layer into the Rio 3.5 model.
Public materials and reports state that Rio 3.5 was produced as a post-train of Alibaba’s Qwen 3.5 397B model. The development has been reported to have cost R$500,000 (approximately $100,000 USD), a figure that Rio did not confirm. Project communications describe the effort as open and as having received public funding from the municipality of Rio de Janeiro. The model is distributed under a fully open MIT license, according to the same disclosures.
This section summarizes the publicly stated background on Rio 3.5’s development, the reported budget figure, the municipality funding claim, and the declared licensing terms. These points reflect the available disclosures and reporting placed alongside the model’s launch information.
Reported benchmark results list Terminal-Bench 2.1 scores of 70.8% for Rio 3.5, 70.3% for Qwen 3.7 Plus, and 67.9% for DeepSeek v4 Pro. Rio 3.5 recorded an IMOAnswerBench score of 89.5%. On the HLE metric, Rio 3.5 scored 36.5%, while Qwen 3.7 Plus scored 34.7%. The numerical comparisons have been published as part of the performance reporting associated with the model’s launch.
Project communications and reporting have touted that Rio 3.5 outperformed leading models on these benchmarks. The comparisons referenced Terminal-Bench 2.1, IMOAnswerBench, and HLE in presenting the model’s performance. This section reproduces the reported benchmark figures without restating promotional claims or endorsements. The Terminal-Bench and HLE comparisons contrast Rio 3.5 directly with Qwen 3.7 Plus and DeepSeek v4 Pro in the cited results.
Eduardo Cavaliere publicly described Rio 3.5 as an open AI model trained in Rio and publicly funded by the municipality, asserting that it had surpassed other models. Public messaging around the launch emphasized local production in Rio de Janeiro and municipal support for the project. Reactions included discourse about ownership and originality, with observers and reporting noting disputes over whether the work represented new original development. Statements and reporting around the launch highlighted both the municipality funding claim and questions about the model’s provenance.
Public and expert reactions thus combined municipal promotion of a Rio-trained, publicly funded model with critical discussion of ownership and originality. Those exchanges framed much of the immediate discourse following the June 13 launch.
