DECK-Arena

DECK-Arena is a benchmark for agentic abilities such as tool use, retrieval and autonomous workflows. The official description says: "Internal Moonshot Elo-score benchmark for presentation generation (python-pptx, pptxjs, in-house DSL; grader calibrated on 10,000+ human judgments)". Below we describe in simple words what benchmarks of this type generally examine.

DECK-Arena measures agentic abilities such as tool use, retrieval and autonomous workflows. In detail, it checks:

• choosing and calling the right tools for a task
• keeping track of what it has already done and what is left
• recovering from failed or contradicting tool results
• completing a defined workflow end-to-end

The score is an Elo rating, built from head-to-head comparisons between models (like chess ratings). The number is only meaningful relative to the other rated models — higher is better.

Official methodology note: Internal Moonshot Elo-score benchmark for presentation generation (python-pptx, pptxjs, in-house DSL; grader calibrated on 10,000+ human judgments).

What is the DECK-Arena benchmark?

DECK-Arena is a benchmark for agentic abilities such as tool use, retrieval and autonomous workflows. The official description says: 'Internal Moonshot Elo-score benchmark for presentation generation (python-pptx, pptxjs, in-house DSL; grader calibrated on 10,000+ human judgments)'. Below we describe in simple words what benchmarks of this type generally examine.

What does DECK-Arena test?

DECK-Arena mainly measures agentic abilities such as tool use, retrieval and autonomous workflows. Concretely, it checks: choosing and calling the right tools for a task; keeping track of what it has already done and what is left; recovering from failed or contradicting tool results.

How is the DECK-Arena score calculated?

The score is an Elo rating, built from head-to-head comparisons between models (like chess ratings). The number is only meaningful relative to the other rated models — higher is better.

Where do I get DECK-Arena data and scores?

The official source is https://huggingface.co/moonshotai/Kimi-K3.1-Pro. The official benchmark page/reference is linked on this page. Every score on Open LLM Radar is also available as open data: download the scores JSON (http://open-llm-radar-api:8000/api/v1/benchmarks/deck-arena/scores) or pull it directly from the Open LLM Radar API with any HTTP client.