4-needle variant at 128k context. 8-needle 128k: 85%.

Highest published match ratio at 256k context. Drops to 76% at 1M tokens.

Long-context retrieval. Strong at 128k, drops to 26.3% at 1M tokens.

8-needle variant at 128k context. Solid long-context retrieval for a 31B dense model.

Sharp degradation at extreme context length. Drops from 77% at 128k.