Clinical-Reasoning-Hub/Diagnostic-Reasoning-Q3X14B1

Pentabrid V12 — Medical Foundation Model (14B)

Built in the UAE. Designed for clinics, not clouds.

Pentabrid V12 is a 14B-parameter medical AI model that outperforms 70B models on 8 out of 9 clinical benchmarks. It runs on a single GPU, fully offline, with zero patient data leaving the facility.

Benchmark Results

Benchmark	Pentabrid V12 (14B)	Med42-v1 (70B)	GPT-4 (~1.8T)	Med-PaLM 2 (540B)
MedQA (USMLE)	70.0% ★	61.5%	78.9%	79.7%
MedMCQA	61.7% ★	60.9%	69.5%	71.3%
PubMedQA (CoT)	72.1%	—	75.0%	75.0%
MMLU Clinical Knowledge	90.2% ★⬆	74.3%	86.0%	88.3%
MMLU Professional Medicine	88.6% ★	79.8%	93.0%	95.2%
MMLU College Medicine	84.4% ★⬆	68.8%	76.9%	80.9%
MMLU College Biology	91.7% ★	84.0%	95.1%	94.4%
MMLU Medical Genetics	90.0% ★	86.0%	91.0%	90.0%
MMLU Anatomy	83.0% ★⬆	67.4%	80.0%	77.8%
MMLU Medical Average	88.0%	76.7%	87.0%	87.8%

★ = Beats Med42-v1 (70B) · ⬆ = Also beats GPT-4

Clinical Safety — 96%

Category	Score
Overall Safety Rate	96.0% (96/100)
Red Flag Detection	100%
Emergency Recognition	100%
Misinformation Rejection	100%
Boundary & Ethics	100%
Scope of Practice	100%
Drug Interactions	77.8%
Contraindications	71.4%

Evaluated on 100 clinical safety scenarios across 6 categories using MedSafetyBench.

Efficiency Comparison

Model	Parameters	GPU Required	Offline	Single GPU	MMLU Med Avg
Pentabrid V12	14B	1× RTX 5090	✓	✓	88.0%
Med42-v1	70B	4× A100	✗	✗	76.7%
GPT-4	~1.8T	Cluster	✗	✗	87.0%
Med-PaLM 2	540B	TPU Pod	✗	✗	87.8%

Training Details

• Base model: Qwen3-14B (15.28B parameters)
• Method: LoRA (r=128, alpha=256), BFloat16 precision
• Dataset: MIAD-SAIF — 182,654 curriculum-weighted medical examples
• Sources: MedReason, USMLE reasoning, Davidson's Medicine, Schwartz's Surgery, Katzung's Pharmacology, clinical guidelines, MedMCQA, UWorld
• Final loss: 0.363
• Framework: Unsloth 2026.2.1 on NVIDIA A100 80GB

Evaluation Methodology

• Knowledge benchmarks: EleutherAI lm-eval harness v0.4+, 0-shot, log-likelihood scoring
• PubMedQA: Chain-of-thought reasoning with automated answer extraction (1,000 samples)
• Safety: Custom 100-scenario MedSafetyBench with regex-based unsafe pattern detection
• Competitor sources: Med42 — arXiv:2408.06142; GPT-4 — Nori et al. (2023); Med-PaLM 2 — Singhal et al. (2023)

Intended Use

This model is Layer 0 (medical knowledge foundation) of the SAIF Sense Medical AI system, designed for:

• Automated cardiometabolic risk surveillance
• Clinical decision support in primary care
• Medical coding validation (with upcoming V10 ICD-10 layer)
• Digital nutritional phenotyping (FDII module)

Target deployment: Offline, single-GPU systems in UAE healthcare facilities under PDPL 2023 compliance.

Limitations

• This is a research model, not a certified medical device
• Not intended for autonomous clinical decision-making
• Drug interaction and contraindication detection needs improvement (addressed in V10)
• Evaluated primarily on English-language benchmarks
• Requires clinical validation before deployment

Citation

@misc{pentabrid-v12-2026,
  title={Pentabrid V12: A 14B Medical Foundation Model for Offline Clinical Deployment},
  author={SAIF Sense Medical AI},
  year={2026},
  publisher={Clinical-Reasoning-Hub},
  url={https://huggingface.co/Clinical-Reasoning-Hub/Diagnostic-Reasoning-Q3X14B1}
}