| Safetensors Size | 66.97 GiB | 18.78 GiB | 13.15 GiB (14,121,080,536 B) | 5.09× compression (−53.82 GiB vs. BF16, −5.63 GiB vs. W4A16) |
| Text Perplexity (378 prompts, 842k tokens) ↓ | 10.8033 | 10.2627 (−5.00%) | 10.3577 (−4.12%) | Chat, math, science, coding, tool use, multilingual |
| Text Divergence vs. BF16 (weighted KL / token mean KL) ↓ | 0.0000 | 0.1139 / 0.2590 | 0.2682 / 0.4338 | Full-vocabulary FP32 KL divergence |
| General & Multilingual Holdout Perplexity (76 prompts, 20k tokens) ↓ | 249.8651 | 257.9844 (+3.25%) | 239.2615 (−4.24%) | Token mean KL 0.4744 (weighted KL 0.5918) |
| Native MTP (Speculative Drafter) (11,583 positions) | 100.0% Top-1 (NLL 4.4308) | 91.44% Top-1 (KL 0.0215) | 91.44% Top-1 (NLL 4.4458) | Isolated MTP KL 0.0215 (end-to-end Top-1 63.88%, KL 0.5776) |
| Vision (VQA, Charts, Documents & OCR) (96 images, 302 tokens) | 1.0000 cos sim · 100.0% Top-1 | 96.36% Top-1 | 95.36% Top-1 | Isolated ViT KL 0.0483 (end-to-end Top-1 92.38%, KL 0.1917) |
| Vision (UI / ScreenSpot & Math OCR) (48 images, 135 tokens) | 1.0000 cos sim · 100.0% Top-1 | 97.04% Top-1 | 97.78% Top-1 | Isolated ViT KL 0.0231 (end-to-end Top-1 86.67%, KL 0.2902) |