MolecularIQ - Counting: leaderboard
Measures symbolic reasoning over molecular graphs in the MolecularIQ paper snapshot: feature counting, atom/bond indexing and constrained molecule generation. The paper reports average accuracy over three rollouts, not pass@3. Chemistry fine-tunes, dated checkpoints and reasoning effort remain distinct. One source Qwen-2.5 14B checkpoint is unspecified and is explicitly labelled rather than assigned to a known base/instruct model. This series reports MolecularIQ - Counting using Mean Accuracy (%).
Metric: Mean Accuracy (%). Source: arxiv.org. Saturation forecast: Around February 2027. 38 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-OSS-120B (High) | 46.8 | #346 (GPT-OSS-120B) |
| 2 | GPT-OSS-20B (High) | 39.1 | #523 (GPT-OSS-20B) |
| 3 | Qwen 3 235B A22B 2507 FP8 (Thinking) | 37.1 | #253 |
| 4 | Qwen 3 Next 80B A3B (Thinking) | 30.7 | #312 |
| 5 | GPT-OSS-120B (Medium) | 29.1 | #346 (GPT-OSS-120B) |
| 6 | GPT-OSS-20B (Medium) | 27.9 | #523 (GPT-OSS-20B) |
| 7 | Seed-OSS-36B-Instruct | 26.8 | #384 |
| 8 | Qwen 3 30B A3B 2507 (Thinking) | 23 | #379 |
| 9 | GPT-OSS-120B (Low) | 17.1 | #346 (GPT-OSS-120B) |
| 10 | Qwen 3 32B | 17 | #449 |
| 11 | GLM-4.6 FP8 | 15.9 | #269 |
| 12 | Qwen 3 14B | 14.4 | #542 |
| 13 | GPT-OSS-20B (Low) | 13 | #523 (GPT-OSS-20B) |
| 14 | Qwen 3 8B | 12.5 | #678 |
| 15 | NVIDIA Nemotron Nano 9B V2 | 12.2 | #855 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=moleculariq-counting · How It Works · Data refreshed daily, snapshot 2026-11-05.