SAFIM: leaderboard
Syntax-Aware Fill-in-the-Middle benchmark evaluating code-completion quality at structurally meaningful insertion points.
Metric: Average. Source: huggingface.co. Saturation forecast: Estimated already saturated. 15 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | deepseek-coder-6.7B-base | 63.4 | #1484 |
| 2 | GPT-4 Preview (1106) | 53.28 | #418 |
| 3 | CodeLlama-34B-hf | 49.66 | #1515 |
| 4 | CodeLlama-7B-hf | 45 | #1457 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=safim · How It Works · Data refreshed daily, snapshot 2026-10-20.