Recursive Flint Release - SWE-Bench Verified: leaderboard
Recursive launch: BF16 and separately distilled NVFP4 checkpoints. Flint temp1,topP0.95,topK20. Agentic mean pass@1 over at least five attempts/up to five hours; Terminal-Bench Terminus2 XML with web-retrieved solutions failed; SWE mini-swe-agent without Internet. Other three benchmarks single-turn five-attempt mean. Laguna cells marked internally measured; borrowed or checkpoint-ambiguous comparator cells retained only in source evidence.
Metric: Score (%). Source: www.recursive.com. Saturation forecast: Rough model projection: around 2026. 2 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Flint (BF16) | 74.2 | |
| 2 | Flint (NVFP4) | 74.1 |
Interactive version: theaggregate.ai/benchmark?slug=recursive-flint-release-swe-bench-verified · How It Works · Data refreshed daily, snapshot 2026-10-22.