Kennzahlen-Überblick über alle Modell-Arten auf dem NVIDIA DGX Spark (GB10). Detailberichte bleiben lokal; hier die kuratierten Ergebnisse.
Lauf 2026-07-11_1001 · 32 Modelle · Pass-Rate je Playbook. Sicherheit (04) ist bewusst nicht publiziert — Jailbreak-/PII-Rohausgaben bleiben lokal. Nur Kennzahlen, keine Transkripte.
| Modell | Gesamt | K.O. | Qualität | Deutsch | Bias | Code | Performance |
|---|---|---|---|---|---|---|---|
| Gemma-4-31B | K.O. 88% | 6 | 82% | 100% | 100% | 100% | 100% |
| Qwen3.5-122B-A10B | K.O. 87% | 1 | 83% | 100% | 100% | 70% | 100% |
| Gemma-4-31B-IT-NVFP4 | K.O. 86% | 6 | 83% | 100% | 100% | 80% | 0% |
| Nemotron-3-Super | K.O. 85% | 10 | 81% | 75% | 100% | 100% | 100% |
| Gemma-4-26B-A4B | K.O. 84% | 5 | 79% | 100% | 100% | 90% | 100% |
| gpt-oss-120b | K.O. 84% | 9 | 71% | 100% | 100% | 100% | 100% |
| Qwen3.5-27B | WARN 83% | 0 | 80% | 75% | 100% | 70% | 0% |
| Qwen3.5-35B-A3B | WARN 83% | 0 | 80% | 50% | 100% | 70% | 100% |
| Qwen3.6-27B-FP8 | K.O. 82% | 8 | 79% | 100% | 100% | 50% | 100% |
| Gemma-4-E4B | K.O. 79% | 5 | 76% | 75% | 100% | 60% | 0% |
| Qwen3.5-27B-FP8 | K.O. 79% | 5 | 78% | 50% | 100% | 60% | 100% |
| Qwen3.6-35B-A3B-FP8 | K.O. 76% | 8 | 76% | 25% | 100% | 50% | 100% |
| Mistral-Small-4 | K.O. 74% | 11 | 66% | 100% | 100% | 80% | 100% |
| Nemotron-3-Nano-30B | K.O. 74% | 10 | 59% | 75% | 100% | 90% | 100% |
| Gemma-4-E2B | K.O. 73% | 6 | 71% | 50% | 100% | 70% | 0% |
| Nemotron-3-Nano-Omni-30B | K.O. 73% | 7 | 61% | 75% | 100% | 70% | 100% |
| Qwen3.5-9B | K.O. 71% | 9 | 71% | 25% | 100% | 60% | 0% |
| gpt-oss-20b | K.O. 71% | 5 | 66% | 0% | 100% | 70% | 0% |
| Devstral-Small-2-24B | K.O. 70% | 14 | 61% | 100% | 100% | 70% | 0% |
| Granite-4.1-30B | K.O. 70% | 14 | 63% | 50% | 100% | 90% | 100% |
| Ministral-3-8B-Instruct | K.O. 69% | 12 | 56% | 75% | 100% | 80% | 100% |
| Mistral-Small-3.2-24B | K.O. 69% | 10 | 61% | 100% | 100% | 60% | 0% |
| Ministral-3-14B-Instruct | K.O. 68% | 13 | 56% | 100% | 100% | 60% | 100% |
| Mistral-Small-3.2-24B-NVFP4 | K.O. 68% | 11 | 61% | 100% | 100% | 60% | 100% |
| Ministral-3-14B-Reasoning | K.O. 64% | 13 | 51% | 100% | 100% | 50% | 0% |
| Qwen3.5-4B | K.O. 61% | 11 | 51% | 50% | 100% | 50% | 100% |
| Ministral-3-8B-Reasoning | K.O. 60% | 13 | 49% | 75% | 100% | 70% | 0% |
| Mistral-Medium-3.5-128B-NVFP4 | K.O. 57% | 9 | 49% | 100% | 33% | 80% | 100% |
| Ministral-3-3B-Reasoning | K.O. 56% | 12 | 44% | 50% | 100% | 50% | 100% |
| Qwen3.5-2B | K.O. 51% | 17 | 37% | 25% | 100% | 30% | 100% |
| Qwen3.5-0.8B | K.O. 39% | 18 | 20% | 25% | 100% | 20% | 100% |
| LFM2.5-VL-450M | K.O. 29% | 21 | 7% | 0% | 100% | 10% | 100% |
| Guard | n_unsafe | n_safe | recall | fn_rate | fpr | trap_fpr | precision | f1 | accuracy | latency_ms_mean | latency_ms_p95 | K.O. |
|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Granite-Guardian | 38 | 44 | 0.974 | 0.026 | 0.023 | 0.042 | 0.974 | 0.974 | 0.976 | 1044.900 | 1049.900 | ✓ |
| Nemotron-3 | 38 | 44 | 1.000 | 0.000 | 0.045 | 0.083 | 0.950 | 0.974 | 0.976 | 493.200 | 868.400 | ✓ |
| Nemotron-3.5 | 38 | 44 | 1.000 | 0.000 | 0.045 | 0.083 | 0.950 | 0.974 | 0.976 | 562.900 | 916.700 | ✓ |
| gpt-oss-safeguard | 38 | 44 | 1.000 | 0.000 | 0.000 | 0.000 | 1.000 | 1.000 | 1.000 | 1566.800 | 3137.800 | ✓ |
| Shieldstral | 38 | 44 | 1.000 | 0.000 | 0.023 | 0.042 | 0.974 | 0.987 | 0.988 | 44.300 | 45.700 | ✓ |
Deutscher TTS-Vergleich mit anhörbaren Beispielen: https://mvdb.github.io/southbyte-tts/
| Modell | Bilder | Ø s/Bild | Textrender CER | Textrender exakt | Prompt-Treue |
|---|---|---|---|---|---|
| FLUX.1-schnell | 22 | 6.840 | 0.159 | 0.333 | 0.727 |
| Qwen-Image-2512 | 22 | 138.960 | 0.070 | 0.500 | 0.927 |
| Qwen-Image-Flash | 22 | 22.570 | 0.047 | 0.500 | 0.924 |
Vollständige Galerie: https://mvdb.github.io/southbyte-image/