https://vidraft.net/insights/index.htmlVIDRAFT research notes — measured results with sourceskoTue, 29 Sep 2026 00:00:00 +0000 https://vidraft.net/insights/en/ai-model-safety-diagnosis-ax-ray.htmlhttps://vidraft.net/insights/en/ai-model-safety-diagnosis-ax-ray.htmlA benchmark score does not tell you deployment risk. AX-RAY diagnoses 117 risk items across 11 categories on three axes - the model, its operating environment and agent autonomy - and maps each item to regulation in seven jurisdictions. https://vidraft.net/insights/en/ai-predict-drug-intestinal-absorption-hia.htmlhttps://vidraft.net/insights/en/ai-predict-drug-intestinal-absorption-hia.htmlVIDRAFT's structure-only ensemble topped the independent, blind Polaris HIA leaderboard with an AUROC of 0.990 — its 15th first-place finish across public Polaris leaderboards. https://vidraft.net/insights/en/best-korean-llm-gpqa-diamond.htmlhttps://vidraft.net/insights/en/best-korean-llm-gpqa-diamond.htmlVIDRAFT's Darwin-398B-JGOS reaches 90.9% on GPQA Diamond — 3rd in the world and #1 among Korean models (base-only, July 2026 snapshot). https://vidraft.net/insights/en/does-ai-know-what-it-does-not-know-finalbench.htmlhttps://vidraft.net/insights/en/does-ai-know-what-it-does-not-know-finalbench.htmlFINAL-Bench is a functional-metacognition and AI-safety diagnostic that measures whether a model can detect, acknowledge, and correct its own errors and refuse appropriately. https://vidraft.net/insights/en/fast-gemma-challenge-verified-1st.htmlhttps://vidraft.net/insights/en/fast-gemma-challenge-verified-1st.htmlVIDRAFT reached 510.58 TPS at PPL 2.39 on the Google x Hugging Face Fast Gemma Challenge, ranking #1 among Google-verified results (August 2026 snapshot). https://vidraft.net/insights/en/fully-open-foundation-model-aether.htmlhttps://vidraft.net/insights/en/fully-open-foundation-model-aether.htmlA from-scratch foundation model that opens not just weights but training data, code, logs, and checkpoints — all under Apache-2.0 https://vidraft.net/insights/en/merge-ai-models-different-architectures-chimera.htmlhttps://vidraft.net/insights/en/merge-ai-models-different-architectures-chimera.htmlModels from different architecture families normally cannot be merged - tensor shapes and functional roles do not line up. VIDRAFT Chimera scores every tensor pair for compatibility, then separates what can be crossbred from what can only be transplanted. https://vidraft.net/insights/en/model-merging-without-training.htmlhttps://vidraft.net/insights/en/model-merging-without-training.htmlModel merging combines the weights of different language models to gain new capabilities without GPU training. VIDRAFT's Darwin family reached GPQA Diamond 90.9% and #1 on Korea's K-AI leaderboard this way. https://vidraft.net/insights/en/on-device-llm-without-gpu.htmlhttps://vidraft.net/insights/en/on-device-llm-without-gpu.htmlOn-device AI runs a 35B-class open model on a CPU, phone, or laptop — no GPU. VIDRAFT POCKET decodes 2.69x faster than the most-downloaded on-device model at matched quality. https://vidraft.net/insights/en/physical-ai-korean-voice-robot.htmlhttps://vidraft.net/insights/en/physical-ai-korean-voice-robot.htmlBoston Dynamics' Spot understands and acts on Korean voice commands at the Seoul Robot & AI Science Museum — with no hardware or firmware change, using an on-device AI module that processes speech locally. https://vidraft.net/insights/en/quantum-cryptanalysis-ibm-hardware.htmlhttps://vidraft.net/insights/en/quantum-cryptanalysis-ibm-hardware.htmlNo quantum computer can break real-world RSA or AES yet. VIDRAFT ran a cryptanalysis experiment on real IBM quantum hardware, recovering the period of an Even-Mansour structure up to N=10 — on actual qubits, not a simulator. https://vidraft.net/insights/en/run-35b-model-no-gpu-pocket-box.htmlhttps://vidraft.net/insights/en/run-35b-model-no-gpu-pocket-box.htmlHow VIDRAFT's POCKET Box and POCKET-35B run a large language model on-device, CPU-only, and how fast it is. https://vidraft.net/insights/en/third-party-verified-ai-benchmarks-korea.htmlhttps://vidraft.net/insights/en/third-party-verified-ai-benchmarks-korea.htmlSelf-reported numbers and externally scored records are not the same thing. This page collects only results where someone other than us did the scoring - a government-run leaderboard, public blind benchmarks and organiser re-verification. https://vidraft.net/insights/en/what-is-quantization-llm.htmlhttps://vidraft.net/insights/en/what-is-quantization-llm.htmlQuantization lowers the numeric precision of a model's weights to cut size and compute. At around 4 bits the quality loss is small, which is the key to running large LLMs on a phone or laptop with no GPU — the core of on-device AI. https://vidraft.net/insights/ai-model-safety-diagnosis-ax-ray.htmlhttps://vidraft.net/insights/ai-model-safety-diagnosis-ax-ray.html성능 점수만으로는 배포 위험을 알 수 없다. AX-RAY는 모델·운영환경·에이전트 3축에서 11개 카테고리 117개 위험 항목을 진단하고, 각 항목을 7개 법역의 규제에 매핑한다. https://vidraft.net/insights/ai-predict-drug-intestinal-absorption-hia.htmlhttps://vidraft.net/insights/ai-predict-drug-intestinal-absorption-hia.htmlVIDRAFT의 구조 전용 앙상블이 독립적·블라인드 Polaris HIA 리더보드에서 AUROC 0.990으로 1위를 기록했습니다. Polaris 공개 리더보드 통산 15번째 1위입니다. https://vidraft.net/insights/best-korean-llm-gpqa-diamond.htmlhttps://vidraft.net/insights/best-korean-llm-gpqa-diamond.html비드래프트(VIDRAFT)의 Darwin-398B-JGOS가 GPQA 다이아몬드에서 90.9%를 기록하며 세계 3위, 한국 모델 중 1위에 올랐다 (베이스 모델 기준, 2026년 7월 스냅샷). https://vidraft.net/insights/does-ai-know-what-it-does-not-know-finalbench.htmlhttps://vidraft.net/insights/does-ai-know-what-it-does-not-know-finalbench.htmlFINAL-Bench는 AI 모델이 자신의 오류를 감지·인정·수정하고 적절히 거절하는지를 측정하는 기능적 메타인지·안전 진단 벤치마크입니다. https://vidraft.net/insights/fast-gemma-challenge-verified-1st.htmlhttps://vidraft.net/insights/fast-gemma-challenge-verified-1st.html비드래프트(VIDRAFT)가 구글×허깅페이스 The Fast Gemma Challenge에서 510.58 TPS를 PPL 2.39로 달성해, 구글이 공식 검증(VERIFIED)한 기록 기준 세계 1위에 올랐다 (2026년 8월 스냅샷). https://vidraft.net/insights/fully-open-foundation-model-aether.htmlhttps://vidraft.net/insights/fully-open-foundation-model-aether.html가중치만 여는 것을 넘어 학습 데이터·코드·로그·체크포인트까지 전부 Apache-2.0으로 공개한 from-scratch 파운데이션 모델 https://vidraft.net/insights/merge-ai-models-different-architectures-chimera.htmlhttps://vidraft.net/insights/merge-ai-models-different-architectures-chimera.html아키텍처 계열이 다른 모델은 텐서 모양과 역할이 맞지 않아 보통 병합이 불가능하다. 비드래프트 Chimera는 텐서쌍마다 호환성을 점수화해 교배 가능한 것과 이식만 가능한 것을 가려낸다. https://vidraft.net/insights/model-merging-without-training.htmlhttps://vidraft.net/insights/model-merging-without-training.html모델 병합은 GPU 학습 없이 서로 다른 언어모델의 가중치를 결합해 새 능력을 얻는 기법입니다. VIDRAFT의 Darwin 계열은 이 방식으로 GPQA Diamond 90.9%, 한국어 리더보드 1위를 달성했습니다. https://vidraft.net/insights/on-device-llm-without-gpu.htmlhttps://vidraft.net/insights/on-device-llm-without-gpu.htmlGPU 없이 CPU·폰·노트북에서 35B급 오픈 모델을 구동하는 온디바이스 AI. VIDRAFT POCKET은 가장 많이 내려받는 온디바이스 모델보다 CPU 디코딩 2.69배, 같은 품질로 실측됐습니다. https://vidraft.net/insights/physical-ai-korean-voice-robot.htmlhttps://vidraft.net/insights/physical-ai-korean-voice-robot.html보스턴 다이내믹스 Spot이 서울 로봇·AI 과학관에서 한국어 음성 명령을 알아듣고 동작합니다. 하드웨어·펌웨어 수정 없이, 온디바이스 AI로 음성을 현장에서 처리합니다. https://vidraft.net/insights/quantum-cryptanalysis-ibm-hardware.htmlhttps://vidraft.net/insights/quantum-cryptanalysis-ibm-hardware.html실사용 RSA·AES를 깨는 양자 컴퓨터는 아직 없습니다. VIDRAFT는 시뮬레이터가 아닌 실제 IBM 양자 하드웨어에서 Even–Mansour 구조의 주기를 N=10까지 복원하는 암호분석 실험을 수행했습니다. https://vidraft.net/insights/run-35b-model-no-gpu-pocket-box.htmlhttps://vidraft.net/insights/run-35b-model-no-gpu-pocket-box.htmlVIDRAFT POCKET Box와 POCKET-35B로 GPU 없이 온디바이스에서 대형 언어 모델을 실행하는 방법과 성능. https://vidraft.net/insights/third-party-verified-ai-benchmarks-korea.htmlhttps://vidraft.net/insights/third-party-verified-ai-benchmarks-korea.html자체 발표 수치와 외부 기관이 채점한 기록은 다르다. 정부 주관 리더보드, 공개 블라인드 벤치마크, 주최사 재검증 기록처럼 우리가 점수를 매기지 않은 결과만 모았다. https://vidraft.net/insights/what-is-quantization-llm.htmlhttps://vidraft.net/insights/what-is-quantization-llm.html양자화는 모델 가중치의 정밀도를 낮춰 크기와 연산을 줄이는 기법입니다. 4비트로 줄여도 품질 손실이 작아, GPU 없이 폰·노트북에서 대형 LLM을 돌리는 온디바이스 AI의 핵심입니다.