| Safetensors Size | 4.24 GiB (4,548,221,488 B) | 1.83 GiB (1,968,201,664 B) | 2.31× compression (−2.41 GiB) |
| Text Perplexity (378 prompts, 840k tokens) ↓ | 3.8302 | 3.7784 (−1.35%) | Chat, math, science, coding, tool use, multilingual |
| Text Divergence vs. BF16 (weighted KL / token mean KL) ↓ | 0.0000 | 0.0591 / 0.2038 | Full-vocabulary FP32 KL divergence |
| General & Multilingual Holdout Perplexity (76 prompts, 20k tokens) ↓ | 12.7820 | 12.8746 (+0.72%) | Token mean KL 0.0500 (weighted KL 0.0630) |
| Native MTP (Speculative Drafter) (11,583 positions) | 100.0% Top-1 (NLL 2.5969) | 92.78% Isolated Top-1 (NLL 2.6173) | Isolated MTP KL 0.0201 (end-to-end Top-1 81.64%, KL 0.3566) |
Vision (vision-probe-v2) (96 images, 302 answer tokens) | 100.0% Top-1 | 97.35% Isolated Top-1 · 93.71% E2E Top-1 | Isolated ViT KL 0.0130, E2E KL 0.0737 (vision-guard-v2 Iso Top-1 99.26%, E2E Top-1 95.56%, KL 0.0562) |