Rio de Janeiro's "homegrown" LLM appears to be a merge of an existing model
Rio de Janeiro’s municipal IT department claimed to have released a “homegrown” large language model that outperformed comparable open models, but researchers quickly showed its weights were almost exactly a 60/40 merge of an existing Qwen3.5 fine‑tune (Nex-N2 Pro) and the original Qwen3.5 model. After the finding, the project’s Hugging Face page was updated to acknowledge the merge and blame an “incorrect upload,” prompting debate over misrepresentation of research capabilities, use of public funds, and the broader need for transparency and provenance standards in open‑weight AI releases.
Allegations About the Rio-3.5 Model
- Rio’s municipal IT arm released “Rio-3.5-Open-397B,” presented as a homegrown post-trained model based on Qwen3.5-397B with strong benchmark results, especially in Portuguese.
- A detailed GitHub issue argues the model is actually a simple merge: ~60% Nex-N2 Pro + ~40% base Qwen3.5-397B-A17B, with no evidence of further post-training or distillation.
- The merged model reportedly responds with the source model’s name and reproduces fine-tuned blurbs from Nex, strengthening the “repackaged merge” claim.
Rio Team’s Response and Transparency Concerns
- The Hugging Face model card was later updated to admit it is a merge of Nex and Qwen plus “on-policy distillation,” and to apologize, claiming the wrong checkpoint (pre-distillation) was uploaded.
- Critics doubt this explanation, noting that the allegedly correct model has not (as of the discussion) been uploaded, and that affiliations on the HF page were edited/removed after the controversy.
- Some see this as misrepresentation of lab capabilities and possible misuse of public funds; others argue intent is unclear and want to wait for a new release and third-party verification.
Technical Discussion: Model Merging
- Commenters note that merging works when models share architecture; here, Nex is itself a Qwen3.5 finetune, so linear interpolation of weights is feasible.
- A simple scheme is described: each weight tensor in the merged model is a weighted average of corresponding tensors from the two source models.
- This technique has precedent (“Frankenstein models”) in both language and image models (e.g., Stable Diffusion), with mixed real-world benefits: often modest gains on specific benchmarks, degradation elsewhere.
- Discussion touches on “linear mode connectivity,” robustness of large models to such merges, and the broader practice of surgically editing weights (merging, “abliteration,” etc.).
Political, Ethical, and Meta Reactions
- Some frame this as a “taxpayer-funded scam”; others note ambiguity about funding, though a public official publicly claimed public money was used.
- There is debate over state vs private role in AI: some favor national/local capability for sovereignty; others see it as state overreach in an area better left to industry.
- Several comments generalize this to a wider pattern of hype, over-claiming, and “marketing-first” behavior across the AI industry.
- A side thread criticizes the use of GitHub issues as public call-out venues and laments a perceived decline in HN discussion quality.