Rio de Janeiro Government Releases 397B-Parameter AI Model with Enhanced Implicit Reasoning

ME AI News: According to monitoring by Beating, IplanRIO, a municipal information and planning company under the Rio de Janeiro city government, has open-sourced Rio-3.5-Open-397B on Hugging Face. The model is post-trained based on Qwen 3.5 397B, employs a MoE architecture, has approximately 397 billion total parameters, activates around 17 billion parameters per token, supports up to 1 million context tokens, and is released under the MIT license. The model card reveals that Rio-3.5-Open-397B integrates the SwiReasoning inference framework. SwiReasoning is a training-free inference method that switches between explicit chain-of-thought reasoning and implicit vector-space reasoning based on changes in information entropy. Explicit reasoning generates conclusions as natural language tokens, while implicit reasoning explores multiple pathways within the hidden space to reduce unnecessary textual output. Test results disclosed by the team show that, with implicit reasoning enabled, Rio-3.5-Open-397B achieves a score of 58.1 on SWE-Bench Pro and 89.5 on IMOAnswerBench. For comparison, the original Qwen 3.5 397B scores 50.9 and 80.9 respectively; when only post-trained without enabling implicit reasoning, it scores 54.8 and 84.5. This indicates that implicit reasoning does not double the model’s absolute performance, but nearly doubles the relative improvement over the base model. Compatibility remains a primary limitation: team members on the Hugging Face discussion forum confirmed these public scores were achieved with SwiReasoning enabled. SwiReasoning requires input of probability-weighted continuous “soft embeddings” during inference, which inference engines like llama.cpp—limited to discrete token ID generation—cannot fully support. The team notes that even without implicit reasoning, the model remains significantly stronger than the original Qwen 3.5 397B, but its full capabilities require inference frameworks that support soft embedding inputs. (Source: BlockBeats)