| Safetensors Size | 66.97 GiB (71,903,871,064 B) | 18.78 GiB (20,160,901,544 B) | 3.57× compression (−48.19 GiB) |
| Text Perplexity (378 prompts, 840k tokens) ↓ | 10.8033 | 10.2627 (−5.00%) | Chat, math, science, coding, tool use, multilingual |
| Text Divergence vs. BF16 (weighted KL / token mean KL) ↓ | 0.0000 | 0.1139 / 0.2303 | Full-vocabulary FP32 KL divergence |
| General & Multilingual Holdout Perplexity (76 prompts, 20k tokens) ↓ | 249.8651 | 257.9844 (+3.25%) | Token mean KL 0.1647 (weighted KL 0.2289) |
| Native MTP (Speculative Drafter) (11,583 positions) | 100.0% Top-1 (NLL 4.4308) | 91.44% Top-1 (NLL 4.4458) | Isolated MTP KL 0.0215 (end-to-end Top-1 73.34%, KL 0.2596) |
| Vision (VQA, Charts, Documents & OCR) (96 images, 302 tokens) | 1.0000 cos sim · 100.0% Top-1 | 0.9856 cos sim · 96.36% Top-1 | Isolated ViT KL 0.0305 (end-to-end Top-1 94.37%, KL 0.0875) |
| Vision (UI / ScreenSpot & Math OCR) (48 images, 135 tokens) | 1.0000 cos sim · 100.0% Top-1 | 0.9900 cos sim · 97.04% Top-1 | Isolated ViT KL 0.0162 (end-to-end Top-1 96.30%, KL 0.0411) |