| Safetensors Size | 51.75 GiB (55,563,006,776 B) | 16.14 GiB (17,330,005,464 B) | 3.21× compression (−35.61 GiB) |
| Text Perplexity (378 prompts, 840k tokens) ↓ | 8.5332 | 8.1841 (−4.09%) | Chat, math, science, coding, tool use, multilingual |
| Text Divergence vs. BF16 (weighted KL / token mean KL) ↓ | 0.0000 | 0.1002 / 0.2583 | Full-vocabulary FP32 KL (math 0.0718, chat 0.0223, if 0.0354, science 0.0607, swe 0.7075) |
| General & Multilingual Holdout Perplexity (76 prompts, 23k tokens) ↓ | 61.0805 | 56.6664 (−7.23%) | Token mean KL 0.0530 (weighted KL 0.0739, 9-lang Belebele macro KL 0.0628) |
| 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 89.16%, KL 0.1782) |
Vision (vision-probe-v2) (96 images, 302 answer tokens) | 100.0% Top-1 | 97.68% Isolated Top-1 · 92.05% E2E Top-1 | Isolated ViT KL 0.0056, E2E KL 0.1808 (vision-guard-v2 Iso Top-1 99.26%, E2E Top-1 94.81%, KL 0.0856) |