| Safetensors Size | 51.75 GiB (55,563,006,776 B) | 11.04 GiB (11,856,436,968 B) | 4.69× compression (−40.70 GiB; 10.58 GiB text-only) |
Text Divergence vs. BF16 (probe-v4) (weighted KL / token mean KL) ↓ | 0.0000 | 0.3137 / 0.4597 | 378 prompts, 852k tokens (PPL 8.5332 → 4.2533; math 0.2051, chat 0.2213, if 0.2112, science 0.2150) |
General & Multilingual Holdout (probe-guard) (token mean KL / weighted KL) ↓ | 0.0000 | 0.3835 / 0.4784 | 76 prompts, 23k tokens (PPL 61.0805 → 24.5085; 9-lang Belebele macro KL 0.4647) |
| Native MTP (Speculative Drafter) (12,339 positions) | 100.0% Top-1 (NLL 2.9579) | 96.64% Isolated Top-1 (NLL 2.9813) | Isolated MTP KL 0.0090 (end-to-end Top-1 72.10%, KL 0.7131) |
Vision (vision-probe-v2) (96 images, 302 answer tokens) | 100.0% Top-1 | 97.68% Isolated Top-1 · 86.09% E2E Top-1 | Isolated ViT KL 0.0056, E2E KL 0.5599 (vision-guard-v2 Iso Top-1 99.26%, E2E Top-1 85.19%, KL 0.5783) |