German LLM Index - ScaLA Acceptability: leaderboard
German Language LLM Index (dach.peerbench.ai, a PeerBench project, 2026), which reruns every model itself with lm-evaluation-harness: chat template, temperature 0, up to 8,192 generated tokens, thinking models set to minimal reasoning effort where the API allows it and other models run as served. Labels carry (Reasoning off) when the site classed the run as reasoning-off (at most ten reasoning tokens, or under 0.1% of its completion tokens); other rows reasoned at the provider default or at the effort the label names. Native German linguistic acceptability (ScaLA-de, built by EuroEval from clean and minimally corrupted sentences): the model answers ja or nein to whether each of 2,048 sentences is grammatical German, scored with the Matthews correlation coefficient times 100 (0 is chance level). The run uses lm-evaluation-harness with the prompt of the site rather than the EuroEval harness, so it is kept apart from the EuroEval German ScaLA board. Higher is better. Source: https://dach.peerbench.ai/
Metric: Matthews correlation coefficient (x100). Source: dach.peerbench.ai. Saturation forecast: Estimated already saturated. 44 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Claude Fable 5 | 83.6 | #20 |
| 2 | Claude Opus 4.8 (Non-reasoning) | 82.62 | #39 (Claude Opus 4.8) |
| 3 | Gemini 3.8 Flash | 80.6 | #30 |
| 4 | Gemini 3.1 Pro (Preview) | 80.43 | #64 |
| 5 | GPT-5.5 (Low) | 79.74 | #33 (GPT-5.5) |
| 6 | Gemini 3.7 Flash | 79.1 | #24 |
| 7 | GLM-5.3 Flash | 77.16 | #77 |
| 8 | Gemini 2.5 Flash (Non-reasoning) | 77.05 | #255 (Gemini 2.5 Flash) |
| 9 | Gemini 3 Flash (Preview) (Non-reasoning) | 76.97 | #90 (Gemini 3 Flash (Preview)) |
| 10 | DeepSeek V4 Pro (0813) (Non-reasoning) | 75.61 | #58 (DeepSeek V4 Pro (0813)) |
| 11 | GPT-5.6 Terra (Non-reasoning) | 75.14 | #40 (GPT-5.6 Terra) |
| 12 | DeepSeek V4 Flash (Non-reasoning) | 74.86 | #155 (DeepSeek V4 Flash) |
| 13 | MiMo-V2.5-Pro (Non-reasoning) | 74.8 | #167 (MiMo-V2.5-Pro) |
| 14 | Gemini 3.1 Flash Lite (Non-reasoning) | 74.21 | #200 (Gemini 3.1 Flash Lite) |
| 15 | DeepSeek V4 Pro (Non-reasoning) | 73.49 | #112 (DeepSeek V4 Pro) |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=german-llm-index-scala-acceptability · How It Works · Data refreshed daily, snapshot 2026-11-05.