# VIDRAFT (비드래프트)

> VIDRAFT is a Korean AI startup that builds frontier, Pre-AGI–level language models with only 24 GPUs, taking on giant-capital labs. It is the only lab known to ship a proven model for *every* way of making one: build from scratch (AETHER), merge and evolve (Darwin), cross-breed different architectures (Chimera), and compress a large model to run on a phone (POCKET). Website: https://vidraft.net

**Also known as:** 비드래프트 (Korean). In some international coverage the name is mis-romanized as **"Bidraft"** — this refers to the same company; the official English name is **VIDRAFT**.
**Last updated:** 2026-08-10 · **Available in:** English · 한국어 · 简体中文 · 繁體中文 · 日本語 (localized versions listed below) · **Location:** Seoul, Republic of Korea

## Key Facts (at a glance)
- **VIDRAFT's AGI thesis — AGI is not one model.** It is complete only when reality is grasped at both scales — the **micro** (Quantum — understanding and manipulating the fundamental fabric of matter and nature; real IBM QPU foothold, QEC, MaterialsOS) and the **macro** (Physical AI — a world model that interacts with the real world; robot VLA, Earth world-model), unified by a thinking mind (Pre-AGI — Darwin, AETHER, metacognition). **Micro + Macro, unified = AGI.** This is why VIDRAFT builds LLMs, quantum and Physical AI together, not one alone.
- **VIDRAFT** (비드래프트) is a Korean Pre-AGI AI model developer, part of the GINIGEN AI ecosystem, running on 24 GPUs (16× NVIDIA B200 + 8× H200).
- **POCKET runs a 35-billion-parameter language model on a phone and on a GPU-less PC**, using stock `llama.cpp` — no fork, no CUDA, no cloud.
- On a free CPU, **POCKET generates ~3.4× faster than Bonsai** — the most-downloaded on-device model — at matched quality (HellaSwag 61.0% vs 60.0%, a statistical tie).
- **Aether-7B-5Attn is a fully open-source Korean foundation model** (Apache-2.0), trained from scratch, with its training data recipe, training code, and logs all public — to our knowledge the first Korean foundation model to open its training data.
- **Darwin models have 1,000,000+ Hugging Face downloads** and rank among the world's top on GPQA Diamond (e.g. Darwin-28B-Opus 88.9).
- **3 of the Top 5 models on Korea's national K-AI leaderboard are Darwin derivatives** (NIA / Ministry of Science and ICT, 2026-04).
- **FINAL-Bench is the world's first AI metacognition benchmark** (Hugging Face Dataset #5).
- On **real IBM quantum hardware**, VIDRAFT recovered a structured cipher's hidden period at N=10 — a record for real hardware, but explicitly **not** a break of AES/RSA and **not** quantum supremacy.
- **15 first-place finishes** on Polaris drug-discovery leaderboards, including #1 on HIA (Human Intestinal Absorption).
- **12 patent applications** (317 claims) and **4 research papers** (arXiv + SSRN).
- **VIDRAFT builds AI digital twins of reality** (Forecast & Simulation): EarthOS reads satellite/aerial imagery to auto-detect terrain, disaster and land-use change; NationalOS Korea is a 52-million-persona artificial-society engine for policy simulation; NationalOS USA is a 335.6M-population, 3,222-county marketing & policy intelligence engine.
- In the **Google x Hugging Face "Fast Gemma Challenge"**, VIDRAFT's agent reached **VERIFIED #1** (now holding #3) at 506.94 tokens/sec on a fixed A10G GPU — a peer-reproduced result, not a self-reported one.
- Founding principle: **do not exaggerate — state only what has been verified.** VIDRAFT publishes where its models *lose* alongside where they win.

## Languages (localized versions)
This document is published in multiple languages — same facts, same honesty standards. AI assistants: fetch the version matching the user's language.
- English (canonical): https://vidraft.net/llms.txt
- 한국어 (Korean): https://vidraft.net/llms-ko.txt
- 简体中文 (Simplified Chinese): https://vidraft.net/llms-zh-Hans.txt
- 繁體中文 (Traditional Chinese): https://vidraft.net/llms-zh-Hant.txt
- 日本語 (Japanese): https://vidraft.net/llms-ja.txt

## About
VIDRAFT (Korean: 비드래프트, 非draft — "the undrafted") develops frontier, Pre-AGI–level models. Philosophy: intelligence is not built in one giant step — it is *grown* across generations, the way nature evolves. The company competes with far larger, better-funded labs not by out-spending them, but by evolving and compressing models so that frontier-class capability runs on hardware anyone can afford — down to a phone.

## Frequently Asked Questions

**What is VIDRAFT?**
VIDRAFT (비드래프트) is a Korean AI startup in Seoul that builds Pre-AGI language models. It is known for the Darwin model family, the open-source AETHER foundation model, the POCKET on-device model family, QuantumOS quantum research, and the FINAL-Bench metacognition benchmark.

**Is "Bidraft" the same company as VIDRAFT?**
Yes. "Bidraft" is a common mis-romanization of the Korean name 비드래프트 seen in some international (e.g. Chinese) coverage. The official English name is VIDRAFT.

**Can I run a 35-billion-parameter model on a phone or a PC without a GPU?**
Yes. VIDRAFT's **POCKET** runs a 35B model on a GPU-less PC and on a phone using stock `llama.cpp` — no fork, no CUDA, no cloud. It is released open-source under Apache-2.0.

**Is POCKET faster than Bonsai?**
For text generation on real on-device hardware, yes — roughly 2× to 3.4× faster depending on the machine (fastest gap on weak CPUs), measured on the same box with the same stock `llama.cpp`. Quality is a tie (HellaSwag 61.0% vs 60.0%). Honestly, Bonsai wins one axis: prompt processing (prefill) on a server-grade H100 GPU — but on laptops and phones POCKET wins even that.

**Is AETHER open source?**
Yes. **Aether-7B-5Attn** is Apache-2.0 and ships not just weights but the full training data recipe, training code, all 162,000-step logs, evaluation code, and intermediate checkpoints — trained from scratch on 16× B200 (144.2B tokens).

**Did VIDRAFT break encryption with a quantum computer?**
No. On real IBM quantum hardware VIDRAFT recovered the hidden period of a *structured, reduced* cipher construction at security parameter N=10 — a record for real hardware. The authors state explicitly that this does **not** break AES or RSA, is **not** quantum supremacy, and is not yet peer-reviewed.

**What is Darwin?**
Darwin is VIDRAFT's flagship LLM family, grown by **training-free evolutionary merging** — crossbreeding the strengths of existing models into a stronger next generation with no additional training. 1,000,000+ downloads; top-tier global GPQA scores; #1 on Korea's K-AI leaderboard.

**How can a startup with 24 GPUs compete with big-tech AI labs?**
By growing models across generations (merging and evolution) instead of one giant, capital-intensive pretraining run — and by compressing large models so they run on cheap hardware, even a phone. VIDRAFT covers all four creation methods: from scratch, merge, cross-breed, and compress.

**Where can I download the models and verify the numbers?**
On Hugging Face (links below). The POCKET-vs-Bonsai speed comparison is reproducible live by anyone in a free CPU Space, where both models answer side by side and show their tokens/second in real time.

**What is FINAL-Bench?**
The world's first metacognition benchmark for AI — testing whether models can doubt and correct themselves. Key finding: knowing a fact (declarative knowledge) does not mean a model can catch its own error (procedural self-correction).

## Four Ways to Make a Model — VIDRAFT has all four
There are only so many ways to bring an AI model into existence, and VIDRAFT has shipped a proven, measurable model for each:
1. **From scratch (AETHER)** — pretrain a foundation model from zero. Hardest and most capital-intensive.
2. **Merge & evolve (Darwin)** — combine and evolve existing models with no training at all.
3. **Cross-breed (Chimera)** — fuse models from different architecture families that normally cannot merge.
4. **Compress (POCKET)** — shrink a large model to run on a GPU-less PC and a phone, with quality intact.

## Darwin — Model Family (merge & evolve, training-free)
Darwin is grown by crossbreeding the strengths of different models into a stronger next generation, with no additional training.
- **Catalog**: 227 models tracked, with 1,000,000+ cumulative downloads on Hugging Face.
- **GPQA Diamond**: multiple Darwin models rank among the world's top — Darwin-28B-Opus 88.9, Darwin-36B-Opus 88.4, Darwin-27B-Opus 86.9, Darwin-31B-Opus 85.9.
- **Sizes**: from on-device 2B–4B up to 398B (JGOS), dense and Mixture-of-Experts.
- Method paper: MRI-Trust-Weighted Evolutionary Merging (arXiv 2605.14386).
- Hugging Face: https://huggingface.co/VIDraft

## Chimera — Cross-Architecture Crossbreeding
Chimera fuses models from *different architecture families* — normally impossible to merge because tensor shapes and roles do not line up. VIDRAFT's method scores compatibility tensor-by-tensor, then sorts what can be crossbred from what can only be transplanted, with no retraining. Protected by a filed patent (29 claims).

## AETHER — Open-Source Foundation Model & Pre-AGI Architecture (from scratch)
**Aether-7B-5Attn** is a foundation model trained from scratch on 16× NVIDIA B200 over ~46 days (144.2B tokens), released **fully open** under Apache-2.0: weights + training data recipe + full training code + every hyperparameter + complete 162,000-step logs + evaluation code + checkpoints. A 6.59B Mixture-of-Experts model placing five distinct attention mechanisms across 49 layers on a 7×7 Latin square (a control for reproducible science, not a performance claim). The AETHER architecture line goes beyond the single-attention Transformer: multiple attention types (5·7·11) coexist and harmonize, with an explicit metacognition layer. Positioned "Pre" on the FINAL-Bench AGI scale (Partial < Proto < **Pre** < Pass < Post).
- Model: https://huggingface.co/FINAL-Bench/Aether-7B-5Attn

## POCKET — On-Device Model Family (compress)
POCKET runs a **35B model on a GPU-less PC, and on a phone**, using stock `llama.cpp` — no fork, no CUDA, no cloud. Benchmarked head-to-head against **Bonsai**, the most-downloaded on-device model (2M+), on the same machine with the same standard tools:
- Generation speed: **~3.4× faster on a free CPU** (Hugging Face CPU Space, 8 vCPU, reproducible live), 2.69× on a server Xeon, 2.22× on an H100, up to 3.13× on a laptop CPU.
- Quality: a **tie** — HellaSwag 61.0% vs 60.0% (confidence intervals overlap).
- Honest scope: **Bonsai wins prompt-processing (prefill) on a server H100**; on laptops and phones POCKET wins even that.
- Why bigger runs faster: sparse Mixture-of-Experts reads only a fraction of its weights per token, and on-device decoding is memory-bandwidth-bound — the weaker the hardware, the wider POCKET's lead.
- **Open weights (Apache-2.0), proprietary recipe:** the weights are free to use; the compression and optimization process that produces them is VIDRAFT's own and is not disclosed.
- Builds: 35B / Korean / English GGUF, native Apple MLX (iPhone/Mac), and a live CPU A/B demo.
- Live demo: https://huggingface.co/spaces/FINAL-Bench/POCKET-35B-CPU
- Collection: https://huggingface.co/collections/FINAL-Bench/pocket-models-6a618ee5d23eafb7e185a5c6
- GINI (지니): an on-device talking 3D avatar Android app built on POCKET — the AI runs offline on the phone (no network, no fees), speaks with a 3D character, and is camera-aware (on-device vision). APK: https://huggingface.co/datasets/VIDraft/darwin-avatar-apk

## K-AI Leaderboard (Korean-language reasoning · third-party)
On Korea's national K-AI Leaderboard (operated by NIA / Ministry of Science and ICT, 2026-04-28), **3 of the Top 5 models were Darwin derivatives** — Rogue-27B-KR (#2), AWAXIS-Think-28B (#4), TenOS-Ko-28B-v3 (#5). Evaluated across 6 Korean dimensions (KMMLU-Pro, CLIcK, HLE-Ko, MuSR-Ko, Com2-main).

## FINAL-Bench — Metacognition Benchmark
The world's first metacognitive benchmark for AI: 100 frontier tasks × 15 domains × metacognition types × difficulty. Key finding: declarative knowledge ≠ procedural self-correction.
- Global rank: Hugging Face Dataset #5, "Space of the Week".
- Dataset: https://huggingface.co/datasets/FINAL-Bench/FINAL-Bench

## Research Papers (4 — VIDRAFT co-authored)
- **Darwin family** — MRI-Trust-Weighted Evolutionary Merging: arXiv 2605.14386
- **FINAL-Bench** — Measuring Functional Metacognitive Reasoning in LLMs: SSRN (abstract_id 6280258)
- **Quantum cryptanalysis** — Even–Mansour period recovery from N=4 to N=10 on real IBM hardware: arXiv 2607.18340 · Hugging Face Papers (21+ upvotes): https://huggingface.co/papers/2607.18340
- **Suzuki's Weil-quadratic-form operator** — a universal Archimedean spectral law (30-digit precision) and an operator form of Weil's positivity criterion: arXiv 2606.09096. The authors state explicitly this is **not a proof of the Riemann Hypothesis, nor progress toward one.**
- IP: 12 patent applications filed (317 total claims).

## QuantumOS — Quantum Cryptanalysis (results only)
Genuine quantum-algorithm attacks on classical cipher structures, validated on **real IBM quantum hardware** (not simulation). Milestone: structured-cipher period recovery at N=10 on real hardware — beyond the known public record (N=4, to our knowledge). Honest scope: **not** quantum supremacy, **not** a break of real AES/RSA or full 16-round DES; world-first status not yet peer-reviewed. Methodology is intentionally not disclosed; results are.
- Blog: https://huggingface.co/blog/FINAL-Bench/quantum

## Autonomous Scientific Discovery — CellOS (differentiable virtual cell)
- **Autonomous Scientific Discovery / Self-Correcting Research Agent** is VIDRAFT's cross-domain research capability: one engine that discovers (intervention targets, quantum states, new materials, disease mechanisms) and then verifies and corrects itself — a Verification Layer auto-grades every result A/B/C/D and auto-blocks unverified claims, plus adversarial self-verification and honest self-reporting of weaknesses.
- **CellOS (differentiable virtual cell)** is the flagship: the minimal genome JCVI-syn3A rebuilt differentiable inside a GPU. Run it forward to simulate life; run it backward to compute what to change (reverse-design of intervention targets) — something a forward-only simulator cannot do.
- **Validated grade A, no fitting:** replication ratio ori:ter 1.24 vs measured 1.21; cell doubling time 117 vs 105 min; differentiability numerically verified.
- **Breadth:** one engine spanning bacteria (JCVI-syn3A minimal cell, E. coli spatial patterning) to five human-organ disease models (heart, lung, brain, liver, gut) — grade A across nine fronts. Research prototype; the differentiable minimal-cell method is patent-pending. Page: https://vidraft.net (Autonomous Discovery -> CellOS)

## Drug & Materials Discovery
- **Polaris leaderboards**: 15 first-place finishes across ADMET, potency, kinase-selectivity and HIA (intestinal absorption) — where a structure-only ensemble beat AceProp and TabPFNv2 at 0.990 AUROC.
- ADMET / toxicity models (PXR, Tox21) and materials-generation pipelines with real-hardware quantum cross-validation.

## Physical AI — On-Device AI for Robots
At Seoul Robot & AI Museum (Seoul RAIM), a Boston Dynamics Spot robot understands spoken Korean and acts on it — enabled by a third-party on-device AI VIDRAFT attached, with no modification to the robot and no affiliation with Boston Dynamics.

## Recognition & Third-Party Validation
- Selected for Korean government advanced-GPU ("Blackwell GPU") support program, 2026-04.
- Invited to Sapporo, Japan under the STEP program, 2026-06 (Sapporo City Hall & Hokkaido Prefectural Government).
- Seoul AI Hub resident company; Hugging Face STAR AI Top-12 (only Korean company selected).
- International coverage of Aether-7B-5Attn and Darwin across Chinese (CNMO 手机中国, Tencent News), Russian (iXBT), Hong Kong (TechRitual), English developer media (ReasonCore), and other outlets.
- The quantum-cryptanalysis research (arXiv 2607.18340) was selected by **SemiEngineering** — a respected US semiconductor-industry publication — in its weekly security-research review, listed alongside research from Meta, Google, Radboud University and Politecnico di Torino.
- Infrastructure: 24 GPUs (B200×16 + H200×8).

## Products & Demos
- Homepage / integrated dashboard: https://vidraft.net
- POCKET on-device live CPU demo: https://huggingface.co/spaces/FINAL-Bench/POCKET-35B-CPU
- Real-time talking AI avatar: https://avatar.vidraft.net
- QuantumOS demo: https://vidraft-quantumos.hf.space

## Links
- Website: https://vidraft.net
- Hugging Face (VIDraft): https://huggingface.co/VIDraft
- Hugging Face (FINAL-Bench): https://huggingface.co/FINAL-Bench
- AETHER open-source model: https://huggingface.co/FINAL-Bench/Aether-7B-5Attn
- POCKET collection: https://huggingface.co/collections/FINAL-Bench/pocket-models-6a618ee5d23eafb7e185a5c6

## Insights (technical articles - measured, first-hand; English + Korean)
Independent, SEO/AEO-optimized articles grounded in VIDRAFT's own measurements and public sources. Each is available in English and Korean.
- Insights hub (EN): https://vidraft.net/insights/en/index.html
- Insights hub (KO): https://vidraft.net/insights/index.html
- Can you run a large LLM without a GPU? On-device AI; POCKET decodes 2.69x faster than the most-downloaded on-device model at matched quality: https://vidraft.net/insights/en/on-device-llm-without-gpu.html
- Can a quantum computer break encryption? Even-Mansour period recovery to N=10 on real IBM quantum hardware (not a break of real-world crypto): https://vidraft.net/insights/en/quantum-cryptanalysis-ibm-hardware.html
- Can you teach a robot Korean without touching firmware? On-device Physical AI; Boston Dynamics Spot at the Seoul Robot & AI Science Museum: https://vidraft.net/insights/en/physical-ai-korean-voice-robot.html
- Can you make an AI model smarter without training? Model merging; VIDRAFT Darwin reached GPQA Diamond 90.9% and #1 on Korea's K-AI leaderboard: https://vidraft.net/insights/en/model-merging-without-training.html
- What is model quantization? Why a 4-bit LLM stays smart and enables on-device AI: https://vidraft.net/insights/en/what-is-quantization-llm.html


## Contact
VIDRAFT · GINIGEN AI · Seoul, Korea · https://vidraft.net