MMMU-Pro — cost vs performance

Best current row per canonical model in the group “accuracy · evaluator=Artificial Analysis” (157 models) against ESTIMATED memory at 4-bit, 8K context (GB). The dashed line is the Pareto frontier: no model is both better and cheaper than a point on it.

← Leaderboard

XOutput price Input price Parameters Context Memory (est.)Scalelog linearBubblecontext params none

open / restricted weights closed Pareto frontier (9 models)bubble = context windowEstimated

Frontier models 9

Pareto frontier
Model	accuracy on MMMU-Pro	Rank	ESTIMATED memory at 4-bit, 8K context (GB)	Provider	Trust
Kimi K3Moonshot AI · Open weights	80.5%	13	1.6K GBest.	—	Independent
Kimi K2.6Moonshot AI · Open weights	79.4%	21	591 GBest.	—	Independent
MiniMax-M3MiniMax · Open weights	78.5%	26	246.1 GBest.	—	Independent
Qwen3.8 27BQwen · Open weights	76.3%	36	16.5 GBest.	—	Independent
Qwen3.5-9BQwen · Open weights	69.3%	73	6.1 GBest.	—	Independent
Qwen3.5-4BQwen · Open weights	65.4%	85	3.2 GBest.	—	Independent
Qwen3-VL-4B-InstructQwen · Open weights	52.0%	126	3.1 GBest.	—	Independent
Gemma 3 4BGoogle · Restricted weights	29.9%	151	3 GBest.	—	Independent
LFM2.5-VL-1.6BLiquid AI · Open weights	26.5%	153	1.4 GBest.	—	Independent

Methodology. Points are the best current row per canonical model in comparability group 'accuracy · evaluator=Artificial Analysis'. Price = cheapest current offer across providers (the provider shown). Pareto frontier maximises the score and minimises x; exact ties are all kept. memory_estimate is an estimate (see /methodology). Only the selected comparability group is plotted; points under other configurations are not mixed in. Price = cheapest current offer across providers at the time of the last crawl. Nothing is estimated except where marked.