KronLSTM Study - Recall 2K - NaturalQuestions: leaderboard
KronLSTM (2026): matched research-trained 1.5B models on 100B pretraining tokens, linear and hybrid architectures kept distinct; 2K-context real-world recall tasks. Architecture names do not identify unrelated public Mamba, mLSTM or GDN checkpoints. Accuracy (%); higher is better.
Metric: Accuracy (%). Source: arxiv.org. Saturation forecast: Rough model projection: around 2026. 12 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | KronLSTM Study KronLSTM 1.5B Hybrid (100B tokens) | 18.82 | |
| 2 | KronLSTM Study KDA 1.5B Hybrid (100B tokens) | 18.12 | |
| 3 | KronLSTM Study GDN 1.5B Hybrid (100B tokens) | 17.9 | |
| 4 | KronLSTM Study GDN-2 1.5B Hybrid (100B tokens) | 17.36 | |
| 5 | KronLSTM Study mLSTM 1.5B Hybrid (100B tokens) | 17.29 | |
| 6 | KronLSTM Study Mamba-3 1.5B Hybrid (100B tokens) | 17.26 | |
| 7 | KronLSTM Study GDN 1.5B Linear (100B tokens) | 16.47 | |
| 8 | KronLSTM Study KDA 1.5B Linear (100B tokens) | 16.31 | |
| 9 | KronLSTM Study GDN-2 1.5B Linear (100B tokens) | 15.9 | |
| 10 | KronLSTM Study KronLSTM 1.5B Linear (100B tokens) | 15.46 |
Interactive version: theaggregate.ai/benchmark?slug=kronlstm-study-recall-2k-naturalquestions · How It Works · Data refreshed daily, snapshot 2026-10-21.