LLM-jp4.1 Release Swallow - AIME 2026 @32: leaderboard
LLM-jp Sept28 release comparison on AIME 2026 @32. Exact source-printed Pass@32 (0–1) with explicit low reasoning effort. Published4/4.1 thinking checkpoints and separately identified SFT-only intermediate ablations remain distinct. Pass@k measures at least one correct solution amongk sampled completions; differentk values are separate boards. Source-specific score; no fusion with a different harness or version is implied. Official Sept28 final release comparison. Exact rounded numbers printed in the heatmap; effort explicitly low. Swallow instruction evaluator; @k is pass@k. SFT-only intermediate ablations are separately labelled; never merge with released SFT+DPO checkpoint. No numeric quantization reported. WMT is source-normalized BLEU. Keep source task and sampler distinct from other publisher results. Metric: Pass@32 (0–1). Higher is better. Exact printed figure values; https://llm-jp.nii.ac.jp/wp-content/uploads/2026/09/image-5-1024x862.png
Metric: Pass@32 (0–1). Source: llm-jp.nii.ac.jp. Saturation forecast: Around January 2027. 27 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | llm-jp-4-33B (High) | 0.93 | #328 (llm-jp-4-33B) |
No result here: #3 Claude Opus 5.5, #5 GPT-6 Astra, #7 Gemini 4 Argon.
Interactive version: theaggregate.ai/benchmark?slug=llm-jp4-1-release-swallow-aime-2026-32 · How It Works · Data refreshed daily, snapshot 2026-10-17.