lighteternal/biodecision-v2-4b

VISIONPricing:Input $0.4 / Cached $0.08 / Output $0.8Concurrent Unit Cost:1Model Size:4.5BQuant:BF16Context Size:32kTool Calling:SupportedPublished:Sep 28, 2026License:otherArchitecture:Transformer0.0K Featherless Exclusive Cold

BioDecision-4B v2 by lighteternal is a 4.5 billion parameter Qwen3.5-based System-1 decision model specifically designed for biomedical, pharmaceutical, and clinical trial applications. It processes a source, question, and answer options to return calibrated probabilities for each answer in a single forward pass, without generating text. This model excels at domain-specific decision tasks, outperforming general models like TypeSafe's Jev 1.13 on many biomedical benchmarks.

Loading preview...

BioDecision-4B v2: Specialized Biomedical Decision Model

BioDecision-4B v2 is a 4.5 billion parameter model developed by lighteternal, built upon the Qwen3.5-4B architecture. It functions as a System-1 decision model, providing calibrated probabilities for given answer options in biomedical, pharmaceutical, and clinical trial contexts, rather than generating free-form text. This version significantly improves upon its predecessor, BioDecision v1, with a 75.3% accuracy across 63,432 held-out test decisions, up from 69.4%.

Key Capabilities & Enhancements

  • Biomedical Decision Accuracy: Achieves 73.9% accuracy on 21 domain decision benchmarks, surpassing TypeSafe's Jev 1.13 (65.8%) in 17 datasets.
  • Expanded Data: Retrained on a 1.75 million-decision pool, incorporating 22 new sources covering drug/device safety, regulatory work, clinical data standards, genomics, and diagnosis.
  • Confidence Calibration: Provides highly calibrated confidence scores; when at least 90% confident, it was correct 98.5% of the time on dev decisions.
  • Relation Extraction: Shows improved performance in drug-protein (DrugProt F1 71.7) and drug-drug (DDI-2013 F1 68.6) relation extraction.
  • Speed: Offers efficient inference with a median of 39 ms per decision and 95 decisions/s in batches on A100 GPUs.

Good For

  • Biomedical Decision Support: Ideal for tasks requiring precise, probability-based answers in areas like drug safety, clinical trial design, and genomics.
  • Replacing TypeSafe's System One API: Compatible with the Jev systemone request format, allowing for a drop-in replacement.
  • Research Use: Suitable for academic and research applications in biomedicine, though not validated for patient care due to non-commercial training sources.

Finetunes
2 models