| Safetensors Size | 17.98 GiB (19,306,310,880 B) | 5.54 GiB (5,951,375,352 B) | 3.24× compression (−12.44 GiB) |
Text Perplexity (probe-v4) (378 prompts, 840k tokens) ↓ | 3.5754 | 2.8545 (−20.16%) | Chat, math, science, coding, tool use, multilingual |
Text Divergence vs. BF16 (probe-v4) (weighted KL / token mean KL) ↓ | 0.0000 | 0.2661 / 0.3435 | 378 prompts, 840k tokens (math 0.1048, science 0.1707, if 0.1979, chat 0.2261) |
General & Multilingual Holdout (probe-guard) (76 prompts, 20k tokens) ↓ | 10.5643 | 10.6139 (+0.47%) | Token mean KL 0.3639 (weighted KL 0.4571, 9-lang Belebele KL 0.4559) |
| Native MTP (Speculative Drafter) (11,583 positions) | 100.0% Top-1 (NLL 2.0275) | 94.86% Isolated Top-1 · 77.46% E2E Top-1 | Isolated MTP KL 0.0105, E2E KL 0.4125 |
Vision (vision-probe-v2) (96 images, 302 answer tokens) | 100.0% Top-1 | 99.34% Isolated Top-1 · 90.73% E2E Top-1 | Isolated ViT KL 0.0073, E2E KL 0.1966 (vision-guard-v2 Iso Top-1 98.52%, E2E Top-1 85.93%) |