Rio's 'Homegrown' 397B LLM Exposed as Nex/Qwen Merge
Get the Tech newsletter
Daily tech — startups, AI labs, chips, the launches that shape the next decade. Free.
- Nex-AGI published a GitHub post asserting that prefeitura-rio/Rio-3.5-Open-397B — presented by IplanRIO as an original 397B model — is a direct element-wise merge of Nex with the official Qwen3.5-397B-A17B base at approximately 0.6 Nex / 0.4 Qwen.
- Nex claims it found no evidence of any original training by IplanRIO, and says Rio's weights are, 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.
- Nex reports that when Rio's hard-coded 'You are Rio' system prompt is stripped, the deployed model self-identifies as 'Nex, from Nex-AGI' 79% of the time and as 'Rio' 0% of the time, and even recites Nex-AGI's bespoke backstory word-for-word.
- Nex notes that other finetunes cannot be explained as such interpolations, and frames this as proof of a direct weight merge rather than conventional fine-tuning.
- IplanRIO had presented Rio-3.5-Open-397B as a city-built model, a claim that Hacker News coverage framed around benchmark performance ('beats Qwen3.7') without addressing the model's provenance.
Why it matters: If Nex's analysis holds, a municipal government has branded and deployed a model built almost entirely by recombining a private startup's weights with Alibaba's Qwen — undermining the 'homegrown AI' narrative and any procurement, sovereignty, or public-interest justification attached to the project.
