Multimodal Knowledge Agent (MKA)

An implementation of Google DeepMind's Multimodal Knowledge Agent (MKA) benchmark. It tests whether agents can read visually dense project files, such as architectural drawings, engineering schematics and clinical scans, and produce the deliverables a professional would.

Models: 21
Median: 41.0%
Range: 27.6%–56.6%

Score distribution

Lowest: 27.6%
Median: 41.0%
Highest: 56.6%

Leaderboard

Include deprecated models

Rank	Model	Score
1	Claude Fable 5.2 (Max, Default Fallback) Anthropic	56.6%
2	Gemini 4 Carbon (High) Google	54.3%
3	Claude Opus 5.5 (Max, Default Fallback) Anthropic	53.0%
4	Claude Sonnet 5.5 (Max, Default Fallback) Anthropic	52.0%
5	Gemini 4 Argon (High) Google	51.3%
6	GPT-6 Astra (Max) OpenAI	50.2%
7	GPT-6.1 Sol (Max) OpenAI	47.7%
8	Muse Wonder (Max) Meta	47.5%
9	Kimi K3.1 Pro Kimi	43.9%
10	GPT-6.1 Terra (Max) OpenAI	43.8%
11	Claude Haiku 5.5 (Max) Anthropic	41.0%
12	Grok 4.7 (Xhigh) SpaceXAI	40.1%
13	DeepSeek V4.1 Flash 1017 (Max) DeepSeek	38.5%
14	Qwen3.8 Max (0902) Alibaba	37.4%
15	MiMo-V2.6-Pro Xiaomi	37.3%
16	GPT-6 Luna (Max) OpenAI	36.8%
17	GLM-5.3-Flash Z AI	36.2%
18	Gemini 3.8 Flash (High) Google	35.3%
19	Mistral Large 4 (High) Mistral	35.1%
20	MiniMax-M3.1-Flash MiniMax	33.5%
21	Step 5 Preview StepFun	27.6%

1
Claude Fable 5.2 (Max, Default Fallback)
Anthropic
56.6%
2
Gemini 4 Carbon (High)

Data from Artificial Analysis

Command Palette

Multimodal Knowledge Agent (MKA)

Score distribution

Leaderboard