K2 Horizon 3.7B (2026-10-01): benchmark results
The checkpoint IFM promoted to K2 Horizon 3.7B's main branch on October 1, 2026: the launch model after a short anti-doom-loop stage (Final Token Preference Optimization) that removes runaway repetitive generations. It shares no weight file with the September release. Provider: Institute of Foundation Models. Released 2026-10-01. Access: Open.
Aggregate Rating 1603 ± 26, rank #920 of 10870 rated models, from 19 benchmark results.
Strongest benchmark results
| Benchmark | Score | Metric | Percentile |
|---|---|---|---|
| K2 Horizon FTPO Card - AIME 2026 | 90.8 | Accuracy (%) | 100 |
| K2 Horizon FTPO Card - BFCL v4 | 64.6 | Accuracy (%) | 66.7 |
| K2 Horizon FTPO Card - HMMT Feb 2025 | 87.5 | Accuracy (%) | 66.7 |
| K2 Horizon FTPO Card - HMMT Feb 2026 | 75.2 | Accuracy (%) | 66.7 |
| K2 Horizon FTPO Card - Terminal-Bench 2.1 | 39.7 | Accuracy (%) | 50 |
| LLM Stats Score | 21.2 | LLM Stats Score (conservative rating) | 45.4 |
| ZeroEval GPQA Diamond | 68.9 | GPQA Diamond Score | 38.9 |
| K2 Horizon FTPO Card - AIME 2025 | 89.2 | Accuracy (%) | 33.3 |
| K2 Horizon FTPO Card - GPQA Diamond | 68.9 | Accuracy (%) | 33.3 |
| K2 Horizon FTPO Card - HLE (text) | 14 | Accuracy (%) | 33.3 |
| K2 Horizon FTPO Card - IFBench (loose) | 47.7 | Accuracy (%) | 33.3 |
| K2 Horizon FTPO Card - LiveCodeBench v6 | 64.5 | Accuracy (%) | 33.3 |
Interactive version: theaggregate.ai/model?slug=k2-horizon-3-7b-2026-10-01 · How It Works · Data refreshed daily, snapshot 2026-11-12.