| Safetensors Size | 8.68 GiB (9,319,828,096 B) | 3.19 GiB (3,429,434,952 B) | 2.72× compression (−5.49 GiB) |
| Text Perplexity (378 prompts, 840k tokens) ↓ | 4.0299 | 3.9537 (-1.89%) | Chat, math, science, coding, tool use, multilingual |
| Text Divergence vs. BF16 (weighted KL / token mean KL) ↓ | 0.0000 | 0.0811 / 0.2133 | Full-vocabulary FP32 KL divergence |
| General & Multilingual Holdout Perplexity (76 prompts, 20k tokens) ↓ | 12.2886 | 12.3692 (+0.66%) | Token mean KL 0.0533 (weighted KL 0.0741) |
| Native MTP (Speculative Drafter) (11,583 positions) | 100.0% Top-1 (NLL 2.2029) | 94.26% Top-1 (NLL 2.2149) | Isolated MTP KL 0.0150 (end-to-end Top-1 85.63%, KL 0.1793) |
| Vision (VQA, Charts, Documents & OCR) (96 images, 302 tokens) | 1.0000 cos sim · 100.0% Top-1 | 99.34% Top-1 (NLL 0.4829) | Isolated ViT KL 0.0050 (end-to-end Top-1 92.38%, KL 0.0889) |
| Vision (UI / ScreenSpot & Math OCR) (48 images, 135 tokens) | 1.0000 cos sim · 100.0% Top-1 | 99.26% Top-1 (NLL 0.5449) | Isolated ViT KL 0.0032 (end-to-end Top-1 95.56%, KL 0.0173) |