MEDIC (closed-ended)

Science

MEDIC is M42's clinical evaluation suite. This row is its closed-ended headline: the average over MMLU medical subsets, MMLU-Pro, MedMCQA, MedQA, USMLE and PubMedQA. Those constituents have different option counts and denominators, so the average admits neither a single lattice nor one guessing floor, and it is carried as a mean of means rather than a proportion. Scored under M42's own harness, a different measurement protocol from lab-reported and Artificial Analysis runs of the same underlying tests, so it forms its own factor group instead of pooling with them.

Top models (higher is better)

Model	Score
Llama 3.1 70B Instruct	79.8
Llama 3.3 70B Instruct	79.5
DeepSeek-V3	78.8
Llama 4 Maverick Instruct	77.7
Llama 4 Scout Instruct	77.0
DeepSeek-R1-Distill-Llama-70B	75.7
Mistral Large 2 (Instruct 2407)	75.0
Qwen2.5 72B	75.0
Llama 3 70B Instruct	74.0
Qwen2 72B Instruct	72.6
Qwen3 32B	71.0
DeepSeek-R1-Distill-Qwen-32B	70.6
QwQ-32B-Preview	69.8
Qwen3 14B	69.0
Gemma 3 27B IT	68.6
Qwen3-30B-A3B	68.5
Phi-4	67.5
Llama 3.1 8B Instruct	67.2
Qwen3-235B-A22B	65.7
Llama 3.1 Nemotron 70B Instruct HF	65.6
Llama 3.1 70B	65.6
Aya Expanse 32B	65.5
DeepSeek R1 Distill Qwen 14B	65.0
Llama 3.2 3B Instruct	63.2
Qwen2.5 7B Instruct	60.0
Solar 10.7B Instruct V1.0	59.4
Qwen3 4B	56.8
Qwen3 8B	56.4
Mistral Large 3 675B Instruct 2512	53.2
Kimi K2 Thinking	52.3
Falcon2-11B	51.9
DeepSeek-V3.1	51.6
Yi 1.5 6B Chat	50.5
Llama 3.1 8B	50.1
Qwen2.5 3B Instruct	49.6
Mistral 7B Instruct v0.3	49.5
DeepSeek R1 Distill Llama 8B	48.4
gpt-oss-120b	42.0
Llama 3.2 1B Instruct	39.4
Phi-3.5 Mini Instruct	39.0