Phi-3.5-mini
A small language model in Microsoft's Phi series with approximately 3.8 billion parameters. It demonstrates strong performance among models of comparable size and is suited for resource-constrained on-device inference scenarios.
Timeline (last 90 days)
The framework was applied to six 7-9B instruction-tuned models (three general-purpose: Mistral-7B, Phi-3.5-mini, Llama-3.1-8B; three medically adapted: BioMistral-7B, OpenBioLLM-Llama3-8B, Med42-Llama3-8B), tested on 200 real patient questions from the K-QA benchmark.
Cross-family routing (e.g., SmolLM 360M routed to Phi-3.5-mini) yields an average improvement of +22.0%, while same-family routing yields an average improvement of only +6.8%.
The dataset covers four open-weight models — Qwen2.5-7B, Mistral-7B-v0.3, Phi-3.5-mini, and Qwen2.5-1.5B — with 30 trajectories each.
All Facts (3)
The framework was applied to six 7-9B instruction-tuned models (three general-purpose: Mistral-7B, Phi-3.5-mini, Llama-3.1-8B; three medically adapted: BioMistral-7B, OpenBioLLM-Llama3-8B, Med42-Llama3-8B), tested on 200 real patient questions from the K-QA benchmark.
50%UnverifiedCross-family routing (e.g., SmolLM 360M routed to Phi-3.5-mini) yields an average improvement of +22.0%, while same-family routing yields an average improvement of only +6.8%.
50%UnverifiedThe dataset covers four open-weight models — Qwen2.5-7B, Mistral-7B-v0.3, Phi-3.5-mini, and Qwen2.5-1.5B — with 30 trajectories each.
50%