Reflection Beam Card - SWEBench Multilingual: leaderboard
Final Oct16 instruction-tuned releases. temp=0.7 top_p=0.99; three numeric variants MUST remain separate; no explicit reasoning effort documented for capability table. Public-only split and with/without tools/context management qualifiers remain in board names. Omniscience is an index, not percent. Avoid Oct5 pre-release scores. Use unique table parent IDs on official model-card HTML, validated against complete headers, row labels and widths; this avoids the README dark/light duplicate tables. Metric: Score (%). Higher is better. Source: https://reflection.ai/beam-model-card
Metric: Score (%). Source: reflection.ai. Saturation forecast: Rough model projection: around 2026. 3 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Beam 1.0 (BF16) | 78.3 | |
| 2 | Beam 1.0 (FP8) | 78.1 | |
| 3 | Beam 1.0 (NVFP4) | 77.5 |
Interactive version: theaggregate.ai/benchmark?slug=reflection-beam-card-swebench-multilingual · How It Works · Data refreshed daily, snapshot 2026-10-17.