Atlas

Benchmarks

← All benchmarks

RULER - 64K

General QA · 2024-04-09

RULER average accuracy over its 13 synthetic retrieval, multi-hop tracing, aggregation and question-answering tasks generated at a 64K-token input length. The number of samples per task (RULER's default or a smaller count such as 100) is a run setting recorded in notes; it is a component of the RULER family, whose primary is the 128K length. Higher is better.

Top models (higher is better)

ModelScore
Nemotron 3 Ultra 550B A55B Base95.3
Qwen3.5-4B95.2
Qwen2.5 14B Instruct 1M94.9
Kimi K2 Base93.8
Nemotron 3 Nano 4B91.5
Qwen2.5-Turbo90.8
Qwen2.5 7B Instruct 1M90.4
Qwen2.5 72B Instruct88.5
Llama 3.1 70B Instruct88.4
GPT-4o Mini87.6
Nemotron 3 Nano 30B A3B Base87.5
Nemotron 3.5 Bolt E4B BF1687.0
Gemma 4 E4B IT86.5
Qwen2.5 Instruct 32B85.3
Qwen2.5 14B Instruct82.3
Qwen2.5 7B Instruct74.5
Loading Atlas data…