| Safetensors Size | 17.98 GiB (19,306,310,880 B) | 6.67 GiB (7,161,432,368 B) | 2.70× compression (−11.31 GiB) |
| Text Perplexity (378 prompts, 840k tokens) ↓ | 3.5754 | 3.4999 (−2.11%) | Chat, math, science, coding, tool use, multilingual |
| Text Divergence vs. BF16 (weighted KL / token mean KL) ↓ | 0.0000 | 0.0638 / 0.1900 | Full-vocabulary FP32 KL (math 0.0144, chat 0.0204, if 0.0218, science 0.0311) |
| General & Multilingual Holdout Perplexity (76 prompts, 20k tokens) ↓ | 10.5643 | 10.6466 (+0.78%) | Token mean KL 0.0395 (weighted KL 0.0586, 9-lang Belebele macro KL 0.0525) |
| Native MTP (Speculative Drafter) (11,583 positions) | 100.0% Top-1 (NLL 2.0275) | 94.86% Isolated Top-1 (NLL 2.0408) | Isolated MTP KL 0.0105 (end-to-end Top-1 87.79%, KL 0.1725) |
Vision (vision-probe-v2) (96 images, 302 answer tokens) | 100.0% Top-1 | 99.34% Isolated Top-1 · 97.35% E2E Top-1 | Isolated ViT KL 0.0072, E2E KL 0.0246 (vision-guard-v2 Iso Top-1 98.52%, E2E Top-1 94.07%, KL 0.0305) |