TanitAI/Tanit-Med-8B
Tanit-Med-8B by TanitAI is an 8.2 billion parameter medical reasoning model, fine-tuned from Qwen3-8B, designed for clinical multiple-choice reasoning and medical question answering. It excels on MedAgentsBench standard splits, achieving a 50.7% macro average, significantly outperforming other open 8B medical models. This model is optimized for USMLE-style vignettes and medical exam reasoning, featuring a unique four-stage training process that includes reasoning SFT and DPO.
Loading preview...
Tanit-Med-8B: An 8B Medical Reasoning Model
Tanit-Med-8B, developed by TanitAI, is an 8.2 billion parameter model based on Qwen3-8B, specifically fine-tuned for medical reasoning and question answering. It underwent a four-stage training process: broad medical SFT, reasoning SFT, DPO, and a chain-of-thought (CoT) polish, with a focus on "thinking before answering" using <think> blocks.
Key Capabilities & Performance
- Leading 8B Performance: Achieves a 50.7% macro average on the standard splits of MedAgentsBench, outperforming the next-best 8B peer by over 16 percentage points.
- Enhanced African Medical Context: Shows a significant lead on AfriMedQA (48.9% vs. 32.2% for the strongest baseline), indicating better performance on questions grounded in African clinical practice.
- Reasoning Focus: Incorporates
<think>blocks, always on by default, to facilitate multi-step reasoning, particularly beneficial for compositional benchmarks like MedBullets and MMLU-Pro. - Competitive with Frontier Models: On standard splits, it performs comparably to last-generation small frontier models like GPT-4o-mini and Claude-3.5-Haiku.
- Robust Training: Utilizes a full fine-tune approach with a capped reasoning budget of 4,096 tokens, optimized for medical questions.
Intended Use Cases
- Medical NLP research and benchmark development.
- Medical education tooling and retrieval-augmented clinical QA prototypes.
- As a base for further fine-tuning in medical domains.
Limitations
- Not for Clinical Use: Explicitly not intended for diagnosis, treatment, dosing, or any patient-facing deployment; it is not a medical device.
- English Only: Evaluated solely on English-language medical content.
- Hard Splits Challenge: Performance on MedAgentsBench hard splits is not meaningfully better than other 8B models, indicating a current scale limitation for highly adversarial questions.
- MedXpertQA Unsolved: Struggles with MedXpertQA, a benchmark with up to ten answer options, similar to other 8B models.