Google DeepMind launches Multimodal Knowledge Agent benchmark

Google DeepMind, working with Artificial Analysis, introduced the Multimodal Knowledge Agent (MKA) benchmark on November 10, 2026. MKA evaluates how AI agents reason across visually complex files to complete real-world knowledge-work tasks, and in initial testing Gemini 4 Carbon ranked second on quality among tested models while costing six times less on average than the top-scoring model.

What the benchmark covers

• The benchmark includes 200 private evaluation tasks, with 10 public tasks released on Hugging Face
• Tasks span 10 verticals and 83 occupations, including architects, biologists, underwriters, and data scientists
• Each task reflects 3 to 8 hours of professional work
• Scoring is weighted 75% factual accuracy and 25% visual design, graded by an automated rater that agrees with human expert grading 96% of the time
• Tested models used 30 to 130 turns on average via the open-source agent harness Stirrup

What it means for AI builders

For people building agentic coding and knowledge-work pipelines, MKA adds a yardstick beyond text-only leaderboards, focused on visual, multi-hour professional tasks. The result that Gemini 4 Carbon placed second on quality while costing six times less than the top-scoring model gives builders a concrete cost-versus-quality tradeoff to weigh when choosing a model for agents that must read visually complex files and produce professional-grade output.

Availability

The 10 public MKA tasks are available on Hugging Face, and the evaluation harness used to test models, Stirrup, is open source, so developers can inspect the task format and the agent loop used to grade factual accuracy and visual design.

Sources

• Introducing the Multimodal Knowledge Agent benchmark (Google, November 10, 2026, official)

Written with AI assistance, strictly from the sources listed above. Before it went live, automated checks verified its dates, sources and every language edition.