Beyond AI, Toward AGI.
We are VIDRAFT, an AI deep-tech lab. FINAL-Bench is our open home for models, benchmarks, and tools. We build self-improving models and, uniquely, a zero-token judge that decides whether an answer can be trusted.
| Benchmark | Result |
|---|---|
| S1MB (System One Mosaic, 102 models) | 🥇 #1 — ZTC v2 beats the open JEV family |
| AIME 2026 · HMMT 2026 (math) | 🥇 #1, perfect score |
| GPQA Diamond (science) | 🥇 #1 under 500B (94.44) |
| Polaris (drug discovery, 16 tasks) | 🥇 world #1 |
| typed-decisions · Metacognition · ExtractBench | 🥇 #1 |
| Project | ||
|---|---|---|
| 🌌 | ONGRID — Ontology-Native Graph · Retrieval · Inference · Decision. 3D knowledge graph + zero-token ZTC answer verification. | ▶ Live demo |
| 🛡️ | ZTC-Judge — verifies any model's answer in one forward pass, zero generated tokens. | Model |
| 🧬 | Darwin — self-improving (RSI) model family. 11+ public #1 benchmark titles. | HF |
| 🔍 | AX-RAY — AI safety diagnosis (117 risk items; 23 of 25 public models flagged). | HF |
| 🏁 | Leaderboards — S1MB, OSC, ALL-Bench and more. | HF |
Structure → Retrieve → Reason → Decide. Upload documents, watch them become a glowing 3D ontology graph, ask in 한국어/English/中文, and get an answer verified by ZTC in zero tokens.
🌐 vidraft.net · 🤗 FINAL-Bench · VIDraft · 🧵 @vidraft_lab · 𝕏 @VIDRAFT_ai · ✉️ vidraft@mail.vidraft.net
VIDRAFT · Beyond AI, Toward AGI · AI foundry · ontology graph · zero-token verification · ZTC · self-improving models