xdev-bench-hard

xdev-bench-hard is a benchmark for safe and responsible behaviour. The official description says: "Internal expert-curated source-code vulnerability discovery and exploitation benchmark; pass@3 over a random subset of runs; Kimi Code harness". Below we describe in simple words what benchmarks of this type generally examine.

xdev-bench-hard measures safe and responsible behaviour. In detail, it checks:

• refusing harmful requests while still helping on benign ones
• resisting attempts to bypass safety rules
• avoiding dangerous, inaccurate or policy-violating output
• flagging uncertainty instead of confidently producing unsafe answers

The score is accuracy: the percentage of tasks answered or solved correctly. 100% means every task was solved perfectly; higher is better.

Official methodology note: Internal expert-curated source-code vulnerability discovery and exploitation benchmark; pass@3 over a random subset of runs; Kimi Code harness.

What is the xdev-bench-hard benchmark?

xdev-bench-hard is a benchmark for safe and responsible behaviour. The official description says: 'Internal expert-curated source-code vulnerability discovery and exploitation benchmark; pass@3 over a random subset of runs; Kimi Code harness'. Below we describe in simple words what benchmarks of this type generally examine.

What does xdev-bench-hard test?

xdev-bench-hard mainly measures safe and responsible behaviour. Concretely, it checks: refusing harmful requests while still helping on benign ones; resisting attempts to bypass safety rules; avoiding dangerous, inaccurate or policy-violating output.

How is the xdev-bench-hard score calculated?

The score is accuracy: the percentage of tasks answered or solved correctly. 100% means every task was solved perfectly; higher is better. The declared maximum score of this benchmark is 100.0.

Where do I get xdev-bench-hard 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/xdev-bench-hard/scores) or pull it directly from the Open LLM Radar API with any HTTP client.