| Safetensors Size | 66.97 GiB (71,903,776,776 B) | 18.78 GiB (20,160,901,544 B) | 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, 840k tokens) ↓ | 4.0208 | 3.7522 (−6.68%) | 3.7230 (−7.41%) | Chat, math (0.0775 KL), science (0.0871 KL), coding, tool use, multilingual |
| Text Divergence vs. BF16 (weighted KL / token mean KL) ↓ | 0.0000 | 0.0847 / 0.2103 | 0.2121 / 0.3885 | Full-vocabulary FP32 KL divergence |
| General & Multilingual Holdout Perplexity (76 prompts, 20k tokens) ↓ | 25.7581 | 24.5254 (−4.79%) | 21.2581 (−17.47%) | Token mean KL 0.1943 (weighted KL 0.2080) |
| Native MTP (Speculative Drafter) (11,583 positions) | 100.0% Top-1 (NLL 2.0268) | 95.80% Top-1 (KL 0.0105) | 95.80% Top-1 (NLL 2.0363) | Isolated MTP KL 0.0105 (end-to-end Top-1 83.47%, KL 0.2470) |
| Vision (VQA, Charts, Documents & OCR) (96 images, 302 tokens) | 1.0000 cos sim · 100.0% Top-1 | 0.9860 cos sim · 99.34% Top-1 | 0.9860 cos sim · 99.34% Top-1 | Isolated ViT KL 0.0072 (end-to-end Top-1 91.72%, KL 0.1385) |
| Vision (UI / ScreenSpot & Math OCR) (48 images, 135 tokens) | 1.0000 cos sim · 100.0% Top-1 | 0.9899 cos sim · 98.52% Top-1 | 0.9899 cos sim · 98.52% Top-1 | Isolated ViT KL 0.0048 (end-to-end Top-1 93.33%, KL 0.1300) |