| Safetensors Size | 51.75 GiB (55,563,006,776 B) | 16.14 GiB (17,328,747,912 B) | 3.21× compression (−35.61 GiB) |
| Text Perplexity (378 prompts, 852k tokens) ↓ | 9.6265 | 9.0881 (−5.59%) | Chat, math, science, coding, tool use, multilingual |
| Text Divergence vs. BF16 (weighted KL / token mean KL) ↓ | 0.0000 | 0.1074 / 0.2831 | Full-vocabulary FP32 KL divergence |
| General & Multilingual Holdout Perplexity (76 prompts, 23k tokens) ↓ | 79.6973 | 72.1364 (−9.49%) | Token mean KL 0.0555 (weighted KL 0.0720) |
| Native MTP (Speculative Drafter) (12,339 positions) | 100.0% Top-1 (NLL 3.1061) | 96.43% Top-1 (NLL 3.1229) | Isolated MTP KL 0.0090 (end-to-end Top-1 89.03%, KL 0.2086) |
| Vision (VQA, Charts, Documents & OCR) (96 images, 302 tokens) | 1.0000 cos sim · 100.0% Top-1 | 98.01% Top-1 (NLL 0.5043) | Isolated ViT KL 0.0071 (end-to-end Top-1 91.39%, KL 0.1444) |
| Vision (UI / ScreenSpot & Math OCR) (48 images, 135 tokens) | 1.0000 cos sim · 100.0% Top-1 | 99.26% Top-1 (NLL 0.8749) | Isolated ViT KL 0.0024 (end-to-end Top-1 92.59%, KL 0.1440) |