ChemRxn-Bench - Bare Pi Strict Reaction F1: leaderboard
603figures; exact normalized reactant/product sets with stereochemistry and one-to-one multiset matching. Per-figure macroF1; missing/malformed/tool-error/timeout predictions receive zero extraction credit. Native Direct/BarePi/RxnMiner harness kept separate; one-hour budget and Qwen thinking apply RxnMiner config only.
Metric: Macro F1 (%). Source: arxiv.org. Saturation forecast: Around February 2027. 5 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Qwen3.8-Flash-Next | 59.6 | #76 |
| 2 | Kimi K3 | 57.7 | #35 |
| 3 | DeepSeek V4.1 Flash | 54.6 | #37 |
| 4 | Seed 2.1 Pro | 26.5 | #73 |
| 5 | GLM-5.3 Flash | 15.5 | #77 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=chemrxn-bench-bare-pi-strict-reaction-f1 · How It Works · Data refreshed daily, snapshot 2026-11-05.