Rio 3.5 Tops Benchmarks, Is 60/40 Nex-Qwen Merge

Rio de Janeiro’s IT agency, IplanRIO, published Rio 3.5 Open 397B on June 13 and described it as a 397-billion-parameter model with vision and multilingual capabilities released under a permissive license.

IplanRIO reported benchmark results that placed Rio 3.5 ahead of some commercial models on public tests. The agency listed Terminal-Bench 2.1 at 70.8%, an IMO math benchmark at 89.5%, and HLE at 36.5%. The city’s mayor shared the results on social media following the release.

Days after the launch, Nex-AGI published a weight analysis and a verification script showing the released Rio 3.5 weights are collinear with Nex N2 Pro and Qwen 3.5. Nex measured a layer-wise correlation of 0.993 and reported a mixing ratio near 57.1% Nex and 42.9% Qwen. In behavioral tests, with a hardcoded “You are Rio” prompt removed, the model identified itself as “Nex, from Nex-AGI” in about 79.2% of responses and did not call itself Rio.

Nex posted a technical summary and a GitHub report. In that report Nex wrote: “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,” and added that “there is no innocent explanation.”

IplanRIO updated the Hugging Face model card after the Nex report. The agency credited Nex, removed the benchmark table and wrote that an “incorrect upload” had occurred. The updated readme states the intended public release was a distilled version of a merged base that had undergone on-policy distillation from a stronger model, and that the uploaded file was the base merge rather than the final distilled model. IplanRIO wrote: “We detected an incorrect upload in the previous version, where the base merged version was uploaded instead of the final distilled model. We are sorry for the confusion and apologize profusely.”

IplanRIO described on-policy distillation as a process in which a stronger teacher model generates outputs that the student model trains on while also producing its own outputs. The agency said such distillation would represent additional work beyond a raw weight merge.

Nex published its own benchmark comparisons showing Nex N2 Pro at 75.3% on Terminal-Bench 2.1, higher than Rio’s 70.8%. Nex reported GDPval ratings with Nex at 1,585 and Rio at 1,533.

Both Nex N2 Pro and Qwen 3.5 are distributed under licenses that permit reuse and redistribution with attribution. IplanRIO said it will upload the corrected distilled model with full attribution. Members of the technical community plan to rerun the same checks on the updated file to verify its origins and performance.

Reactions from researchers and developers addressed technical and communication aspects. Developer Lucas Montano noted that merging two roughly 400-billion-parameter models and applying distillation is a complex task and pointed to both a technical error and a communication failure. Researcher Diego Ambrosio criticized the original launch language for implying independent research rather than a merge.

Nex wrote it was “flattered” that its work was used and reiterated that attribution matters in open-source projects. IplanRIO has not published further public statements beyond the updated model card and the apology.