MedXpertQA (MM): published results

TsinghuaC3I (Tsinghua University) · 2,000 multimodal questions (MM subset) · index updated September 28, 2026

GPT-5.6 Sol has the highest indexed numerical score on MedXpertQA (MM), 81.5 as of , per Qwen3.8-Max: A New Bar for Coding and Cowork. Expert-level multimodal medical multiple-choice QA covering clinical images (X-ray, histology, dermatology, charts) across 17 specialties; MM subset of the 4,460-question MedXpertQA benchmark.

Published results

#	model		score	as of
1		GPT-5.6 Sol OpenAI Qwen-run comparison in the Qwen3.8-Max launch post; Primary blog currently renders an empty shell; retained historical value, not reverified on 2026-09-28. via Qwen3.8-Max: A New Bar for Coding and Cowork · partial review	81.5	not reported
2		Gemini 3.1 Pro Google Meta's Muse Spark launch table (Meta reports the better of vendor self-reports and its own reproduction) via launch post, ai.meta.com	81.3%	2026-04
3	A	Qwen3.8 Max Alibaba Alibaba's own Qwen3.8 launch table; protocol not stated; Primary blog currently renders an empty shell; retained historical value, not reverified on 2026-09-28. via Qwen3.8-Max: A New Bar for Coding and Cowork · partial review	80.4%	2026-08
4		Claude Fable 5 Anthropic Qwen-run comparison in the Qwen3.8-Max launch post; Primary blog currently renders an empty shell; retained historical value, not reverified on 2026-09-28. via Qwen3.8-Max: A New Bar for Coding and Cowork · partial review	80.0	not reported
5		Muse Spark Meta Meta's Muse Spark launch table (Meta reports the better of vendor self-reports and its own reproduction) via launch post, ai.meta.com	78.4%	2026-04
6		GPT-5.4 OpenAI Meta's Muse Spark launch table (Meta reports the better of vendor self-reports and its own reproduction) via launch post, ai.meta.com	77.1%	2026-04
7		Gemini 3 Pro Google Qwen3.5 model card comparison; source-specific evaluation, not harmonized across vendors via Qwen/Qwen3.5-397B-A17B model card	76.0%
8		GPT-5.2 OpenAI Qwen-run comparison in the Qwen3.5-397B-A17B model card via Qwen/Qwen3.5-397B-A17B model card	73.3	not reported
9		Claude Opus 4.8 Anthropic Qwen-run comparison in the Qwen3.8-Max launch post; Primary blog currently renders an empty shell; retained historical value, not reverified on 2026-09-28. via Qwen3.8-Max: A New Bar for Coding and Cowork · partial review	71.7	not reported
10	A	Qwen3.7 Plus Alibaba Alibaba's own Qwen3.7 Plus launch table; protocol not stated; Primary blog currently renders an empty shell; retained historical value, not reverified on 2026-09-28. via Qwen3.7-Plus: Multimodal Agent Intelligence · partial review	71.0%	2026-05
11	A	Qwen3.5 397B A17B Alibaba self-reported in the Qwen3.5-397B-A17B model card via Qwen/Qwen3.5-397B-A17B model card	70.0	not reported
12	A	Qwen3.6 Plus Alibaba Qwen-run comparison in the Qwen3.7-Plus launch post; Primary blog currently renders an empty shell; retained historical value, not reverified on 2026-09-28. via Qwen3.7-Plus: Multimodal Agent Intelligence · partial review	68.7	not reported
13		Grok 4.20 xAI Meta's Muse Spark launch table (Meta reports the better of vendor self-reports and its own reproduction) via launch post, ai.meta.com	65.8%	2026-04
14		Kimi K2.5 Moonshot AI Qwen-run comparison in the Qwen3.5-397B-A17B model card (K2.5-1T-A32B column) via Qwen/Qwen3.5-397B-A17B model card	65.3	not reported
15		Claude Opus 4.6 Anthropic Meta's Muse Spark launch table (Meta reports the better of vendor self-reports and its own reproduction) via launch post, ai.meta.com	64.8%	2026-04
16		Claude Opus 4.5 Anthropic Qwen3.5 model card comparison; source-specific evaluation, not harmonized across vendors via Qwen/Qwen3.5-397B-A17B model card	63.6%
17		Gemma 4 31B Google Google model card; MedXPertQA MM row; vendor-reported; protocol differs from other source families via Gemma 4 model card	61.3%
18		Gemma 4 26B A4B Google Google model card; MedXPertQA MM row; vendor-reported; protocol differs from other source families via Gemma 4 model card	58.1%
19		Gemma 4 12B Google Google's Gemma 4 model card, Unified 12B; protocol not stated via Gemma 4 model card	48.7%	2026-04
20	A	Qwen3-VL-235B-A22B Alibaba Qwen3.5 model card comparison; source-specific evaluation, not harmonized across vendors via Qwen/Qwen3.5-397B-A17B model card	47.6%
21		Gemma 4 E4B Google Google model card; MedXPertQA MM row; vendor-reported; protocol differs from other source families via Gemma 4 model card	28.7%
22		Gemma 4 E2B Google Google model card; MedXPertQA MM row; vendor-reported; protocol differs from other source families via Gemma 4 model card	23.5%

Scores preserve their source precision, with any scale conversion documented (mixed sources). Assembled table, not a single board. Rows come from several vendors' own launch documents: Meta's Muse Spark evaluation (whose values live only in a rendered table image and cover Meta's own model plus four competitors it re-ran or quoted), Alibaba's Qwen3.5, 3.7 and 3.8 posts (Alibaba's own runs of its models and of competitors), and Google's Gemma 4 model card. Protocols are not known to match across rows; each row's source line says which document it came from and who ran it. benchlm.ai presents a subset as one mirrored view while citing only Meta's methodology. Checked 2026-09-28: Meta image table visually verified from its methodology PDF; Qwen3.5 and Google model-card tables verified. Six Qwen3.7/3.8 blog rows retain their historical retrieval dates and partial confidence because those primary pages currently return no article content. The line under each model names the document its score was read from; rows marked source pending are still awaiting a documented first-party source. Full citations are on the sources page.

About the benchmark

The rest of the field is on the index, and how sources qualify is on the methodology page. Model names in the table link to cross-benchmark pages.