| Safetensors Size | 1.63 GiB (1,746,942,600 B) | 0.79 GiB (844,779,168 B) | 2.07× compression (−0.84 GiB) |
| Text Perplexity (378 prompts, 840k tokens) ↓ | 4.1695 | 4.3015 (+3.17%) | Chat, math, science, coding, tool use, multilingual |
| Text Divergence vs. BF16 (weighted KL / token mean KL) ↓ | 0.0000 | 0.0735 / 0.2833 | Full-vocabulary FP32 KL (math 0.0271, science 0.0370, if 0.0392, chat 0.0471) |
| General & Multilingual Holdout Perplexity (76 prompts, 20k tokens) ↓ | 15.4384 | 16.3150 (+5.68%) | Token mean KL 0.0706 (weighted KL 0.0887, no_robots KL 0.0554) |
| Native MTP (Speculative Drafter) (11,583 positions) | 100.0% Top-1 (NLL 2.8764) | 90.24% Isolated Top-1 (NLL 2.8997) | Isolated MTP KL 0.0288 (end-to-end Top-1 78.36%, KL 0.4515) |
Vision (vision-probe-v2) (96 images, 302 answer tokens) | 100.0% Top-1 | 94.04% Isolated Top-1 · 88.08% E2E Top-1 | Isolated ViT KL 0.0266, E2E KL 0.1388 (vision-guard-v2 Iso Top-1 97.04%, E2E Top-1 94.81%, KL 0.0884) |