UCSC-VLAA/ClinSeek-35B-A3B
ClinSeek-35B-A3B is a 35.1 billion parameter model developed by UCSC-VLAA, fine-tuned from Qwen/Qwen3.5-35B-A3B with a 32768 token context length. It specializes in agentic clinical reasoning by learning multi-round evidence-seeking processes from raw Electronic Health Records (EHR) using native tool calls. This model significantly outperforms its base model and other open-source alternatives on the AgentEHR-Bench, demonstrating advanced capabilities in retrieving and reasoning with patient-specific clinical evidence.
Loading preview...
ClinSeek-35B-A3B: Agentic Clinical Reasoning Model
ClinSeek-35B-A3B is a 35.1 billion parameter model developed by UCSC-VLAA, derived from Qwen/Qwen3.5-35B-A3B. It is specifically fine-tuned to automate multimodal evidence seeking for agentic clinical reasoning, learning from trajectories generated by GPT-5.6 Sol. Unlike traditional models, ClinSeek-35B-A3B is designed to actively retrieve patient-specific evidence from raw Electronic Health Records (EHR) using integrated tools before formulating answers.
Key Capabilities
- Agentic Clinical Reasoning: Learns a multi-round evidence-seeking process, reasoning within its native thinking block and executing actions via native tool calls.
- Superior Performance: Achieves an average F1 score of 35.6 on AgentEHR-Bench, marking a +16.1 improvement over its base model. It surpasses all evaluated open-source models, including larger ones like Qwen3.5-397B-A17B, and even closed-source models like Claude Opus 5.
- Enhanced Evidence Retrieval: Demonstrates significant improvements in precision (18.6% to 32.4%) and recall (31.1% to 53.3%) compared to its base model. It effectively writes complex SQL queries (53% of questions vs. 0.8% for base), reads more available tables (72% vs. 36%), and utilizes imaging or discharge notes more frequently (74% vs. 12%).
- Tool Integration: Designed to work with ClinSeekAgent EHR tools and a multi-turn tool loop, enabling dynamic interaction with clinical data.
Good for
- Automated Clinical Evidence Seeking: Ideal for applications requiring models to actively query and interpret raw EHR data.
- Clinical Decision Support Research: A strong foundation for developing and evaluating agent-based systems in healthcare.
- Benchmarking Agentic Models: Provides a high-performing open-source baseline for tasks involving complex clinical reasoning and tool use.
This model is a research tool and should not be used for clinical diagnosis, treatment, or patient management.