Atlas

Benchmarks

← All benchmarks

Arena-Hard-Auto v2.0 - Hard Prompts and Creative Writing Average

Chat & Writing · 2025-04-23

The unweighted mean of the Arena-Hard-Auto v2.0 hard-prompts and creative-writing win rates, as NVIDIA's Nemotron releases report it under 'Arena-Hard-V2 (Average)'. A component of the Arena-Hard v2.0 family (arena-hard-v2), computed from the same runs as its two subsets. Higher is better.

Top models (higher is better)

ModelScore
Nemotron Cascade 2 30B A3B83.5
Qwen3.5 35B A3B65.4
Qwen3 30B A3B Thinking 250757.8
gpt-oss-20b48.5
Loading Atlas data…