| Safetensors Size | 17.98 GiB (19,306,303,864 B) | 6.67 GiB (7,157,574,568 B) | 2.70× compression (−11.31 GiB) |
| Text Perplexity (378 prompts, 843k tokens) ↓ | 6.6829 | 5.6579 (-15.34%) | Chat, math, science, coding, tool use, multilingual |
| Text Divergence vs. BF16 (weighted KL / token mean KL) ↓ | 0.0000 | 0.1440 / 0.5751 | Full-vocabulary FP32 KL divergence |
| General & Multilingual Holdout Perplexity (76 prompts, 20k tokens) ↓ | 15.0115 | 14.8568 (-1.03%) | Token mean KL 0.0965 (weighted KL 0.1336) |
| Native MTP (Speculative Drafter) (11,583 positions) | 100.0% Top-1 (NLL 2.8335) | 94.68% Top-1 (NLL 2.8539) | Isolated MTP KL 0.0127 (end-to-end Top-1 79.17%, KL 0.4402) |
| Vision (VQA, Charts, Documents & OCR) (96 images, 302 tokens) | 1.0000 cos sim · 100.0% Top-1 | 97.02% Top-1 (NLL 0.5313) | Isolated ViT KL 0.0205 (end-to-end Top-1 94.70%, KL 0.1109) |
| Vision (UI / ScreenSpot & Math OCR) (48 images, 135 tokens) | 1.0000 cos sim · 100.0% Top-1 | 97.04% Top-1 (NLL 0.9373) | Isolated ViT KL 0.0107 (end-to-end Top-1 94.81%, KL 0.0741) |