DeepMind Unveils Multimodal Knowledge Agent Benchmark
Google DeepMind and Artificial Analysis introduced a visual, multi-step benchmark for evaluating AI agents on professional tasks.
Updated on Nov. 10, 2026 in Artificial Intelligence
Google DeepMind and Artificial Analysis released the Multimodal Knowledge Agent (MKA) benchmark on November 10, 2026. The benchmark tests AI agents on 200 private tasks across 10 industry verticals that combine text and visual reasoning.
Why it matters
Many existing AI benchmarks rely on text-only tasks, and MKA is designed to add visual reasoning to evaluation. Whether it becomes an industry standard depends on how other labs respond and whether they submit models.
MKA tasks were shaped with input from specialists across 83 occupations and are modeled on projects that typically take human professionals 3 to 8 hours. Scoring uses an agentic autorater that matched human expert judgments 96.2% of the time, though the tasks themselves remain private.
The players
Google DeepMind
Google DeepMind is the artificial intelligence research laboratory of Google that develops advanced AI models and evaluation methods.
Artificial Analysis
Artificial Analysis is an organization that independently benchmarks and evaluates artificial intelligence models.
The details
Agents must interpret architectural blueprints, engineering schematics, and clinical scans, and produce deliverables such as spreadsheets, slide decks, and marked-up images. Scoring weighs factual accuracy alongside visual design.
Timeline
November 10, 2026: DeepMind and Artificial Analysis unveil MKA with a blog post and paper.
The Tech Race
MKA extends the trajectory of agent evaluation from text-only tests toward multimodal professional work. Its release positions DeepMind and Artificial Analysis to shape how labs measure agent performance against rival evaluation efforts.
Professionals who rely on spreadsheets, slide decks, and visual documents may see AI agents become more capable at these tasks as evaluations like MKA push labs to improve. Gemini 4 Carbon's lower cost per completed task could also make agent tools less expensive to use.
The takeaway
Benchmarks that combine visual reasoning with professional workflows give a fuller picture of what AI agents can actually do. Readers following AI tools should watch whether other labs adopt MKA-style testing.
Further reading
Explore more on Artificial Intelligence developments and evaluation methods.
