Building a Pre-AGI platform in Korea, aiming to become a scientific research company — Darwin delivers results, AETHER designs the future.
Everything VIDRAFT builds sits on one of three axes — and each one feeds the other two.
Our own architecture, trained from scratch — the work of moving from Pre-AGI toward AGI. AETHER places multiple sequence-mixing mechanisms on a Latin square so the question stays answerable.
Quantum compute applied where GPUs struggle — search and optimisation. Two aims: securing high-value substance IP in drugs and materials, and the research that carries us toward the AGI stage.
Intelligence and judgement added to robots through an attachable on-device module built on our own models — no platform modification, no cloud dependency.
The axes are not separate businesses. The foundation LLM gives Physical AI its on-device intelligence; QuantumOS post-processes what the models propose; the substance IP it secures funds the next model.
There are only so many ways to bring an AI model into existence: build it from nothing, merge existing ones, fuse different architectures, or compress a big one to run anywhere. VIDRAFT is one of the very few labs that has shipped a proven model with every single one — each backed by a number you can check.
Pretrain a foundation model from zero — the hardest, most capital-intensive path.
Combine and evolve existing models with no training at all.
Fuse models from different architecture families — which normally cannot merge at all.
Shrink a large model to run on a GPU-less PC and a phone — with quality intact.
Key achievements by VIDraft in 2026.
In an editor-curated weekly review, our quantum cryptanalysis paper — run on real IBM quantum hardware — was listed next to papers from Meta, Google, Radboud and Politecnico. Not a press release we sent out, but an editorial selection: our first earned pickup by overseas trade press.
Read →A foundry that builds AI models to order, without pre-training. Applying the semiconductor foundry model to AI: the customer brings the design intent, the foundry runs the process. Platform construction is in progress.
Joint project · platform build in progressDual-engine strategy — VKAE (up to 23.4× throughput) + VKUE (35B frontier models on consumer GPU, notebook and CPU). CEO Kim Min-sik: grow into an infrastructure company that makes big AI cheaper, faster and closer.
Read →180 of 198 correct. World-class performance through pure reasoning. MoE + Darwin V9 method.
Read →MSIT·NIA certified score 0.621, ranked #1. 8 of top 12 are Darwin derivatives (67%).
Read →China’s largest portal features Darwin’s “genetic recombination” merging — smarter AI without retraining.
Read →VIDraft across Korean tech media & the founder's essays.
AGI is achieved step by step. VIDraft is now at the Pre-AGI stage.
Single-task specialization
Multi-task combination
Metacognition + Self-correction + Swarm collaboration
True AGI
Post AGI
Sovereign fully-open (weights · data · training code · logs · checkpoints) — 6 countries side by side.
K-AI #1 · GPQA Diamond 90.9% · Polaris Global 14× Champion · Metacognition Leaderboard #1
Merging pretrained models is cheap; merging them without losing what made each good is not. Three filings, 77 claims.
Blend two models at one uniform ratio and you mix one model’s strength with the other’s weakness. Quantifying each layer’s functional importance sets a different ratio per layer — attention and FFN are never merged on the same setting.
In plain terms: when combining two dishes you do not mix the broth and the solids in the same proportion.
Models from different architecture families have mismatched tensor shapes and roles, so they cannot simply be merged. Scoring compatibility tensor by tensor separates what can be crossbred from what can only be transplanted — with no retraining.
In plain terms: like an organ transplant — you test compatibility first, then decide what can be grafted and what merely attached.
A single merge stops where it stops. Feeding children back as parents accumulates gains across generations, and re-injecting parent tensors breaks the model out of evolutionary stagnation.
In plain terms: keep breeding across generations — and bring in outside stock when the line starts to narrow.
HuggingFace certified · As of June 2026
| Rank | Model | Score | Note |
|---|---|---|---|
| ★ | Darwin-398B-JGOS | 90.9% | VIDraft Latest |
| 1 | Darwin-28B-Opus | 88.89% | VIDraft |
| 2 | Qwen3.5-397B-A17B | 88.40% | Alibaba |
| 3 | Kimi-K2.5 | 87.60% | Moonshot |
| 5 | Darwin-36B-Opus | 88.40% | VIDraft |
| 11 | Darwin-27B-Opus | 86.90% | VIDraft |
| 16 | Darwin-31B-Opus | 85.90% | VIDraft |
| 21 | Darwin-9B-NEG | 84.34% | VIDraft |
MSIT·NIA certified · Blind evaluation · As of June 2026
| Rank | Model | Score | Category |
|---|---|---|---|
| 1 | JGOS-31B-Citizen | 0.621 | VIDraft Official |
| 2 | AWAXIS-KR-31B-v5 | — | Darwin Derivative |
| 6 | Rogue-28B-MIX | — | Darwin Derivative |
| 7 | AWAXIS-Hybrid-28B | — | Darwin Derivative |
| 8 | Warecube-KO-27B-v3 | — | Darwin Derivative |
| 9 | Rogue-27B-KR | — | Darwin Derivative |
| 11 | AWAXIS-Think-28B | — | Darwin Derivative |
| 12 | TenOS-Ko-28B-v3 | — | Darwin Derivative |
A result that emerged naturally in the MSIT·NIA blind evaluation. Third-party certified figures, not self-reported.
Not a one-off merge — a parent → child → grandchild structure across successive generations.
First evolution from base models
An 85% GPQA score at 4B parameters
Cross-architecture breeding (Mamba)
20+ official · 1,200+ derivatives & quant variants · 1M+ all-time downloads (HF-verified 2026-07)
GGUF 양자화 코딩 특화 모델. 최근 30일 22,537 다운로드로 다윈 패밀리 GGUF 최다 — 노트북·로컬에서 바로 구동. (base: Darwin-28B-Coder)
9B NEG single-pass evolution. Most-downloaded model in the Darwin family — 575K+ all-time.
GPQA Diamond 88.4%. Community-favorite flagship, widely quantized and redistributed.
35B MoE flagship. GPQA Diamond 90.0%. Multimodal support.
Korean flagship. CLIcK 84.5 · KMMLU 76. Top Korean-language Darwin.
Coding specialist. Top official GGUF download in the Darwin family.
Transformer × Mamba cross-breeding. CLIcK 92%. World-first heterogeneous crossbreed verified.
GPQA Diamond 88.89%. Surpasses 400B class with no extra training. arXiv 2605.14386.
Reasoning-enhanced specialist. Optimized for math·code·science reasoning.
31B-class at 4.5B. GPQA 85.0%. Cumulative +26.4%p. Fit for embodied AI & mobile.
Public administration Korean. CLIcK 0.987 · KMMLU-Pro 0.725 · Com2 0.742.
27B BF16 legal specialist. B200 load 9.7s. TPS 22.3 tok/s.
1.7B Korean speech synthesis. Cross-architecture evolution.
MoE structure, 17B active params. Darwin V9 method. World-class pure reasoning.
⬇️ = HuggingFace all-time downloads (official + community redistributions), verified 2026-07. Darwin family total: 1M+.
Browse and download all 20 models in the HuggingFace FINAL-Bench collection.
Adaptive Multi-way Triple-symmetric Hybrid attEntion Routing
Putting several attention mechanisms in one model is not the hard part. Making it actually train is. Placement skews, gradients diverge per type, causality leaks, experts collapse. These eight filings block those failures in order — 224 claims across the stack.
Mix attention types carelessly and you cannot attribute a good result — was it the composition, or where each type happened to land? Fixing each type to appear exactly once per row and column makes the layout deterministic and reproducible.
In plain terms: spreading tasting booths evenly across every floor, so you can say the menu sold well — not just the third floor.
Summing outputs from different attentions lets the loudest one dominate. A single learnable gate fuses them — and the gate itself is regularised so it cannot collapse onto one mechanism.
In plain terms: one mixing fader, with a lock that stops any single channel from going to 100%.
When future tokens leak into past computation, training looks great and deployment fails. Previously this meant auditing layers by hand. Here the leak is located automatically and the offending module is swapped for a causally safe one.
In plain terms: catching a student peeking at the answer page — and replacing the method, not just the score.
Fixing it once is not enough — the next commit can reintroduce it. Any model or code change triggers re-verification, and a failure blocks the merge.
In plain terms: safety enforced by the pipeline, not by someone remembering to check.
Each attention type carries different gradient statistics; one global clip destabilises training. Measuring each type separately catches anomalies before divergence, not after.
In plain terms: you do not load the same weight onto athletes of different size.
MoE routinely collapses onto a few experts while the rest go unused. Instead of coaxing balance with auxiliary losses, a prime-based symmetric layout secures it structurally.
In plain terms: rather than nagging people to take turns, you write a rota that turns by construction.
Different inputs demand different kinds of reasoning, yet experts are usually fixed. Here the model judges what reasoning an input needs, selects accordingly, and can grow new reasoning types during inference.
In plain terms: sorting the question first — "this is arithmetic, this is common sense" — then calling the right person.
Moving to a model of different shape usually means training from scratch. Cyclic-group structure decides which layers correspond, and weights move only within matching attention types and functional groups.
In plain terms: when you move house, kitchen things go to the kitchen and books to the study — you do not tip every box onto the floor.
가중치만 던지는 것은 컴파일된 바이너리를 주는 것과 같다. Aether-7B-5Attn은 데이터 레시피·토크나이저·학습 코드·모든 하이퍼파라미터·전체 로그·평가 코드·중간 체크포인트까지 전부 Apache-2.0으로 공개했다. 성능 우위를 주장하려는 것이 아니라, 처음부터 다시 만들 수 있는 '진짜 오픈'과 주권을 증명하려는 것이다.
'완전한 오픈'은 지금까지 국가연구소·정부 컨소시엄·대학 연합의 몫이었다. Aether는 단일 스타트업으로 그 명단에 이름을 올렸고, 유일하게 이종 어텐션(5종)을 얹었다.
O 공개 · △// 부분·제약 · X 없음·미정 | Aether = 6.59B MoE · 어텐션 5종을 7×7 라틴방진에 배치
49개 층(7×7)에 어텐션 라벨을 라틴 스퀘어로 배치하면, 각 메커니즘이 모든 깊이에 정확히 한 번씩 나타난다. 깊이 편향을 제거한 실험 통제 장치다.
※ compress·linear = full 계열 → 5종 구별 메커니즘. NSA 분기는 패딩 마스크를 소비하지 않으므로 추론은 batch_size=1.
Published on Hugging Face — 3 base models + 1 instruct variant.
Fully open — weights · data recipe · training code · 162k-step logs · checkpoints. 49 layers (7×7 Latin square) · 6.59B total / ~2.98B active (MoE) · Apache-2.0
Instruction-tuned variant of the open-source base. Same architecture · Apache-2.0
Open weights only — architecture and training details not disclosed. 6.59B total / ~3B active (MoE) · base stage · Apache-2.0 weights
121 layers (11×11 Latin square) · 11 sequence-mixing mechanisms across 4 families · 5.79B (MoE) · mid-training checkpoint (step 174,000) · Apache-2.0 weights
Steps 110k / 115k / 162k — released for reproduction.
Aether Sovereign AI — try the model in your browser.
All Aether models in one place.
Planned for release. Specifications are subject to change.
4-way attention · 34B class
4-way attention · 70B class
MoE · 34B total / 10B active class
MoE · 460B total / 30B active class
Most production models run one attention mechanism in every layer. That is a convention, not a proven optimum.
| Model | Attention composition | Structural limitation |
|---|---|---|
| Llama-3 | Full · 1 type | 80 layers · rapid KV-cache growth |
| Mistral | Sliding window · 1 type | Limited long-range dependency |
| Mamba2 | State space · 1 type | No softmax · accuracy trade-off |
| DeepSeek-V3 | MLA + Full · 2 types | Strong, but limited attention diversity |
| AETHER | Multiple types on an n×n Latin square | Placement is an experimental control — superiority still to be established by ablation |
Integrating the latest research into a single LLM.
DeepSeek
Microsoft
Vaswani 2017
DeepSeek-V2
Longformer
Native Sparse Attention · Sequence compression
Differential Attention · Adaptive noise cancellation
ICML 2020
Gu & Dao 2024
Delta Network
Poli ICML 2023
Teaching a model to sense when it might be wrong — a research effort to reduce hallucination.
A model that recognizes "I might be wrong" and re-checks. Our goal: make honesty and self-correction measurable, not assumed.
On an internal Korean hallucination test, a 5-way ensemble raised the rate of catching its own errors from 33% to 67% — a research result, not a shipped product.
We built FINAL Bench, a benchmark that scores whether a model knows what it does not know — because you cannot improve what you cannot measure.
49 layers (7×7 Latin square) · weights, data recipe, training code, logs and checkpoints all published · Apache-2.0
Instruction-tuned variant of the open-source base
7-way and 11-way heterogeneous stacks — 11Attn is a mid-training checkpoint (step 174,000)
Scale-up line. Specifications are subject to change.
9 emergence engines + MARL runtime
8-type self-error detection · TICOS
Closed-loop perpetual learning · SLAI
Multi-agent · SOMA/MOUSE
Five-element enhance/suppress · Hallucination control
JinigenAI · VIDraft joint development · 24 AI models evaluated · Published 2026.07.01
JGOS-31B-Citizen avoids 398 of 400 trap questions (99.5%, #1). Metacognition = an essential AGI capability.
World's first metacognition benchmark. 100 tasks × 15 domains × 8 types (TICOS) × 3 difficulty levels.
400 questions designed to induce plausible-but-wrong answers. Measures self-error recognition.
Analyzes internal signals to detect self-errors in advance without altering the original model's structure or weights.
11 individual metacognition adapters for major models like Darwin-28B-Opus, Darwin-9B-NEG, and JGOS-31B-Citizen.
Metacognition performance comparison leaderboard. 24 AI models evaluated across trap avoidance and self-error detection.
An open-source on-device model family that runs a 35B model on a GPU-less PC, and on a phone, with stock llama.cpp — no fork. Measured against Bonsai on the same machine, we published where we win and where we lose.
A Korean startup's on-device model, benchmarked head-to-head against the world's most-downloaded one — on the same machine, with the same stock llama.cpp. We show where we win, and the one row where we lose.
On Hugging Face’s free CPU Space, POCKET and Bonsai answer side by side on the same 8-vCPU box — and each shows its tok/s in real time. POCKET generates ~20.5 tok/s vs Bonsai ~6.1 — about 3.4× faster (CPU-only). Ask anything and watch.
⚔️ Open the live A/B demo →35B total · ~3B active per token · sparse MoE
27B dense · reads every weight per token
most-downloaded on-device model (2M+)
| METRIC | Bonsai | RESULT | |
|---|---|---|---|
| Desktop / Server (Linux) | |||
| ★ HF CPU Space — 8 vCPU (reproducible) | ~20.5 | ~6.1 | 3.4× faster |
| CPU generate — Xeon, 16 threads | 27.0 | 10.1 | 2.69× faster |
| GPU generate — H100 | 197 | 89 | 2.22× faster |
| GPU prompt — H100 | 753 | 1816 | Bonsai wins |
| Quality — HellaSwag, 400q | 61.0% | 60.0% | = tie |
| Laptop — MacBook M3 Pro, 18 GB | |||
| Metal generate | 25.4 | 12.8 | 1.99× faster |
| CPU generate — 8 threads | 13.8 | 4.4 | 3.13× faster |
| Metal prompt | 240.7 | 73.4 | 3.28× faster |
Same machine, same unmodified llama.cpp — anyone can reproduce it. POCKET needs no fork; the comparable Bonsai build will not even load in upstream llama.cpp.
Apache-2.0POCKET is not just an open model re-saved in a smaller file. Its weights are Apache-2.0 — free for anyone — but the compression and optimization process that produces them is VIDRAFT’s own, and it is not published. What is open is the artifact; what is ours is how it is made.
Weights, configs and usage are all Apache-2.0 — download, run, modify, redistribute. No strings.
The recipe that keeps a 35B model coherent at extreme small sizes is VIDRAFT’s proprietary technology — developed in-house, and not disclosed.
Coherent quality even at 8.2 GB, with dedicated builds for Korean and for English — exactly where a naive one-size quantization falls apart.
Not one model, but on-device builds split by device and language.
PC / server, GPU-less CPU. IQ1_M (8.2 GB) to Q4_K_M (21 GB).
Korean build for Android and local runtimes. From 5.1 GB.
Native Apple silicon (MLX) for iPhone and Mac.
English build. PC and iPhone mixes.
Live demo — answers with no GPU, in your browser.
26B (Gemma-based) build. GGUF for GPU-less CPU.
Live 26B demo — no GPU, in your browser.
POCKET is a sparse Mixture-of-Experts model: many experts, but only a few activate per token. On-device decoding is bound by memory bandwidth — how much you read each step — so reading a fraction of the weights makes a larger model faster where there is no GPU. And unlike the comparable Bonsai build, which needs a vendor fork to load, POCKET runs on stock llama.cpp — so LM Studio, Ollama and the standard ecosystem work on day one.
Honest scope: server-GPU (H100) prefill goes to Bonsai; on laptops and phones POCKET wins prefill too. Quality is tied on HellaSwag — neither side claims a quality win. iPhone speed is not yet self-measured (community reports welcome). Real-world feel varies with RAM, storage, quantization and thermals.
Apache-2.0 · open-source on-device model family by VIDRAFT
What we published, and what we filed.
Each benchmark proves a different capability — and each was administered by an external party.
Korean & public-sector LLM capability
PhD-level scientific reasoning
Drug-candidate prediction & screening
Medical hallucination control
Serving optimisation on constrained hardware
Metacognition & self-error detection
Darwin-398B-JGOS 90.9%. Darwin-28B-Opus 88.89% #1. 5 Darwin models in Top 21. HuggingFace certified.
JGOS-31B-Citizen overall #1 (0.621). 11 Darwin derivatives in the Top 20. Blind evaluation.
HF Global Top 5. 100 tasks × 15 domains × 8 types × 3 difficulty. TICOS quantitative measurement.
World model benchmark. Sole model to pass Track C. Real-time interactive world simulation.
Structure-only prediction — no cell-imaging or external experimental data — topped 14 public Polaris leaderboards across absorption, distribution, metabolism, excretion, toxicity, potency and kinase selectivity.
All results are computational predictions on public blind-submission leaderboards, submitted under the VIDraft account. VIDraft has since filed its first drug-substance patent from this autonomous discovery system.
MRI-Trust-Weighted Evolutionary Merging for Training-Free Scaling. 45 days after release: 300K+ downloads, 500 derivatives.
Measuring Functional Metacognitive Reasoning in LLMs. MetaCog scaffold +14.05 pts. 94.8% Error Recovery. Public preprint on SSRN · under conference review.
Extending Even–Mansour period recovery from N = 4 to N = 10 on real IBM quantum hardware (ibm_kingston, Heron). Five genuine attacks across four symmetric-cipher paradigms, validated to the 25-qubit classical-simulation ceiling.
Scope stated explicitly: Q2 query model, reduced/structured constructions, follows the birthday bound — this is not quantum advantage, and does not break full AES/RSA or 16-round DES.
The first numerical realization of Suzuki’s Hilbert–Pólya candidate operator: a universal Archimedean spectral law pinned to 30-digit precision, with a conductor-only offset, and an operator form of Weil’s positivity criterion. The authors state explicitly that this does not prove the Riemann Hypothesis, nor advance toward a proof.
KR priority filing + PCT · 317 claims total
Neural-network parameter merging based on Model-layer Response Importance (MRI)
Compatibility-Based Crossbreeding for Heterogeneous Neural Networks
Recursive Generational Inheritance in Evolutionary Neural Network Merging
5/7/11-way heterogeneous attention Latin Square integration
Causal Safety Verification for Multi-Type Attention
Self-error detection signal head
Single Sigmoid Gate Convex Combination Fusion
Prime-Order Cyclic Group Weight Transfer
Per-Attention-Type Gradient Clipping
Prime-Based Mixture-of-Experts Routing
Causal Safety CI/CD for Neural Network Models
N-(pyrrolidin-1-yl)-thiazol-4-yl boronic acid compound for progressive supranuclear palsy (PSP) — composition, preparation method and use
On-device AI built on our own LLM — added to robots and physical systems without touching the platform.
A complete Physical AI kit — software and hardware — that attaches to a robot without any prior arrangement with its manufacturer.
Vision-language model built on our own LLM — perception and language in one stack.
The control harness that binds model output to robot action and safety limits.
Our inference acceleration and lightweight engines — latency, low-end hardware, throughput.
9 hours of independent operation per charge
We add intelligence and perception to a robot as a third party — no partnership with the manufacturer, no firmware fork, no change to the platform. The same kit that runs on a Boston Dynamics Spot can be fitted to other machines already in the field.
Three constraints that cloud AI cannot solve for physical systems.
A robot acting on a spoken command cannot wait on a round trip. Inference has to happen where the sensor is.
Defence, industrial and public sites often run without outbound network. The model has to work offline.
Systems already deployed will not be rebuilt around AI. Intelligence has to arrive as an attachment.
Boston Dynamics builds the robot. We made it understand Korean — as a third party, without touching their platform.
Spot is a quadruped robot built by Boston Dynamics. As shipped, it does not understand Korean.
No firmware fork, no vendor partnership, no changes to the platform. We attached an on-device AI module.
A visitor speaks Korean. The model on the device interprets intent and the robot moves accordingly.
Deployed as a live exhibit at the Seoul Robot & AI Science Museum — operating for general visitors, not a lab demo.
Most physical systems already deployed in the field will never be rebuilt around AI. If intelligence can be added as an attachable on-device module, the installed base becomes addressable — without the manufacturer's involvement. That is the on-device AI supply model behind our Physical AI line.
A visitor speaks Korean. Our on-device model — mounted on Spot and running on the robot itself with no cloud connection — interprets the intent, and Spot acts on it. Nothing in the robot was modified; the intelligence arrived as an attachment. The whole proposition is in this one clip.
Specifications from the manufacturer. Our kit mounts within the 14 kg payload allowance — no modification to the robot itself.
Three exhibits built by VIDraft, running for the public.
Boston Dynamics 'Spot' controlled by natural-language voice — converse with, command and steer the robot through an LLM.
See inside the LLM black box — an X-RAY view that makes a model's internal workings transparent to visitors.
Watch AI models crossbreed and evolve — the child surpasses both parents. An interactive Darwin evolution experience.
A small Korean Proto-AGI startup (since 2024) with one belief: don't exaggerate — speak only what is verified.
I am Min-sik Kim, CEO of VIDRAFT. I am twenty-eight — a fourth-year computer engineering student at Korea Cyber University, currently on leave. I paused school to build this company, so my credentials will not tell you much.
So let me tell you what I have done instead. I am named inventor on 12 patent applications and co-author on 4 papers. All of that intellectual property belongs to VIDRAFT, by design — it is the company’s foundation, not my résumé.
The AGI axis holds Darwin, Chimera and AETHER. Darwin diagnoses a trained model (Model MRI) and breeds it across generations — three patents: reliability-weighted merging, cross-architecture breeding, recursive generational inheritance. Chimera decomposes a model’s knowledge structure and transplants it, which led to our own architecture, AETHER — eight patents covering heterogeneous attention on a Latin square and metacognitive verification. Eleven of our AI patents sit on this axis.
On quantum computing, I am a co-author on the paper that extended Even–Mansour period recovery on real IBM Heron hardware from the public record of N = 4 to N = 10. On the same axis we filed a substance patent for a progressive supranuclear palsy therapeutic compound — we search with quantum, and capture the value as substance IP.
On Physical AI, we attached an on-device model built on our own LLM to a Boston Dynamics Spot so that it understands and acts on spoken Korean. It runs today at the Seoul Robot & AI Science Museum.
I serve as principal investigator on two government programmes — the AI computing-resource utilisation project, completed and rated Excellent, and an ongoing advanced-GPU support programme.
Before that, I volunteered for signals duty and completed my military service. What I took from it was not how to work the equipment, but a plainer fact: once a connection drops, nothing else functions. The soundest judgement counts for nothing if it never reaches anyone.
And before I was someone who reads research, I was someone who gathered people to read it together. Some 15,000 people are active members of the communities I run — ArXivGPT, OpenFree AI and others — where we share new work every day. Much of VIDRAFT’s ability to absorb research quickly comes from there.
I want to win a Nobel Prize in the sciences — Physics, Chemistry, or Physiology or Medicine. Those are precisely the three domains we work in: quantum for physics, new materials for chemistry, drug discovery for medicine. That alignment is deliberate, not a coincidence.
And one ambition that should outlast me: a Max Planck Society for Korea. Germany’s Max Planck spent a century funding basic research that paid nothing in the short term, and produced more than thirty Nobel laureates. Korea does not yet have a vessel like that. I want this company to become one — earning from products, spending on fundamental research, and letting that research become the next product. That is why our business rests on two axes: solutions and services, and substance IP.
That is a large claim, which is exactly why we say only what has been verified. Please check our published models, papers and patents yourself. Exaggeration gets found out; the record remains.
VIDRAFT Inc. is a Seoul-based deep-tech company that, built on Pre-AGI-level AI and quantum computing, extends its research and development into physics, chemistry, life sciences, and pharmaceuticals.
The Darwin model family and the AETHER architecture — intelligence grown to the Pre-AGI stage.
QuantumOS — verifying quantum cryptanalysis step by step on real hardware, tilling the soil for the next era.
New-material and molecular research — checked first against novelty and prior-art gates.
Drug-candidate and peptide discovery — first drug-substance patent filed in 2026.
Great life was never designed all at once. Seed → sprout → tree → forest. AGI is grown by nature's law, not built by force.
We do not build intelligence in one giant step; we grow it across generations, the way nature evolves great life. 씨앗 → 새싹 → 나무 → 숲.
Our founding spirit: do not exaggerate; speak only what has been verified. Real trust comes from honestly stating what we have NOT done.
Crossbreed the strengths of different models to raise a stronger next generation — inheriting a lineage of intelligence, like a craft passed down through generations.
Not one method — 5·7·11 kinds of attention coexist. When diversity harmonizes, higher intelligence awakens. This is how we look beyond the Transformer.
We do not claim to have 'completed' the quantum computer. We till the soil for the next era, verifying step by step on real hardware. AGI is a road, not a destination.
Technology for people and society — the same dream as Society 5.0. From Jeonnam·Gwangju (3.2M citizens) to across borders.
Eleven people, armed with freedom, flexibility, speed and creativity — like romantic pirates.
Ask any VIDraft member their identity and we answer 'Jo-mark' — the crew's callsign, drawn from the manga One Piece. Not resources but wits; not volume but grit and agility.
낙천이 전략이고, 확신이 자산이다.
회장이 소리쳐도, 이 배의 항로는 흔들리지 않는다.
숫자는 냉정하게, 팀은 진심으로.
이 바닥의 모든 항로를 지도 없이 안다.
명함 한 장에 왕국이 딸려온다.
일본과 세계를 잇는 유일한 사무라이.
우리가 지금 만드는 것들이 인류의 미래가 될 것이다.
Corporate Research Institute, recognised by KOITA (No. 2026111898).
말없이 자리를 지키고, 결과로 답한다.
CEO가 자면 자고, CEO가 코딩하면 코딩한다.
정식 크루가 되는 그날까지, 배우겠습니다.
칭찬하지 마세요! (매우 기뻐하는 중)
For now I keep ciphers. When I return, I break them.
A tight team that beats volume with method — 24 GPUs against the giants.
Own LLMs · City OS · drug/material AI · on-prem AI — CEO, Vice President and Japan Business Head lead a Proto-AGI stack.
Looking for the next global AI pirate — freedom, speed, creativity. Contact us.
Three revenue axes: solutions and services delivered on-premise, substance IP generated by our domain engines, and core technology licensed to partners.
Domain OS delivered as on-premise deployments, PoCs and managed services. Revenue from licences, deployment and operations.
PharmaOS and MaterialOS generate drug and material candidates that become filed intellectual property — an asset that compounds independently of service revenue.
Core architecture and process technology licensed to partners on a non-exclusive basis. The partner runs the platform and owns the customer relationship; we supply the process. First case: the 'Model Foundry' platform being built with Answerwise Inc. (under construction)
These are existing revenue lines, not roadmap items.
Full AI transformation delivered inside the customer network — model, serving stack and workflow. Data never leaves the perimeter.
On-device vision-language models for robots and physical systems. Reference deployment: Seoul Robot & AI Science Museum.
NationalOS / JGOS public-administration engine, plus PharmaOS and MaterialsOS domain builds and managed operations.
Why on-premise: public-sector, defence, healthcare and financial data often cannot leave a secure network — a market general-purpose cloud AI is not well suited to serve.
Market entry through domain-specific OS. Counterparties in ongoing negotiations are not disclosed.
Regional public-administration LLM (JGOS) — development and PoC
Drug-discovery biofoundry — lead role planned
Logistics-specialised AI stack built on partner-provided operational data
VK Engine-based infrastructure business — domestic and overseas
MaterialsOS — collaboration with a state-backed materials body
2026 → 2028 · market validation, vertical OS commercialisation, multi-domain Pre-AGI platform.
NationalOS/JGOS public-administration PoC · PharmaOS·MaterialOS domain validation · VK Engine on-prem packaging
AETHER V1 commercial release · public/medical/materials OS commercialisation · Series A and partner expansion
Global enterprise entry on proprietary architecture · OS commercialisation scale-up · PCT filings and global IP strategy
Not one chatbot, but an AI operating system that connects administration, citizen and industry data safely inside a local / independent network.
Administration overload · fragmented data · cloud-AI security concerns → a local sovereign AI operating system.
Language · administration · industry models tuned for each region.
Hallucination suppression · self-verification · evidence-backed answers.
Policy · budget · industry simulation on regional digital twins.
A shared AI access point for citizens, officials and enterprises.
Vertical-domain AI plugged into the regional OS.
On-premise / independent-network deployment, verified by a 3-month KPI pilot.
From natural language to KS-standard floor plans (DXF·BIM·BOQ·fire-safety), and from a parcel address to code-compliant 3D building massing.
Three sibling engines: VKAE runs it fast (≈9× single-GPU), VKUE runs it anywhere (down to a free CPU), and VKIE serves the most (18,057 tok/s on one B200). Every number measured.
The real AI-infra battleground is not building the model — it is running it, cheaply, forever.
GPUs are scarce, expensive and power-hungry. Doubling throughput on the same card is like building twice the infrastructure — no new chips, no new power.
Training happens once; inference happens every time a user asks. VKAE lowers the per-token cost that decides whether an AI service is profitable.
For clouds, AI startups, enterprises and national/sovereign AI, acceleration is not optional — without it, running AI at scale runs at a loss.
Speed claims are cheap. VKAE ships the model weights plus the optimized serving stack as one Docker container, OpenAI-compatible — so anyone can reproduce the numbers on their own GPU and plug it straight into an existing service.
The ubiquity engine that serves frontier-class models on minimal hardware. VKAE runs fast; VKUE runs everywhere.
Run GPQA 86–88% models on a single 24GB consumer GPU — or on CPU with no GPU. (Scores use different methods; compare with care.)
| Model | GPQA Diamond (method) | VKUE min. HW |
|---|---|---|
| Darwin-398B-JGOS | 90.9 (greedy) | 1 GPU + RAM |
| Darwin-36B-Opus ⭐ | 88.4 (maj@8) | 1× 24GB consumer GPU / CPU |
| Ourbox-35B-JGOS ⭐ | 86.4 (maj@8) · 70.7 (greedy) | 1× 24GB consumer GPU / CPU |
| GLM-5.2 (744B) | frontier-class | 1× H100 · speed under verification |
GPQA 88.4. Runs on a single 24GB consumer GPU or CPU.
GPQA 86.4 (maj@8). Ubiquitous serving on consumer hardware.
Quantized GGUF build — the same weights that run on a laptop or a GPU-less CPU.
The full VKUE explainer — how one file runs from a B200 down to a GPU-less CPU.
A 35B model running live inside a free, HF-provided CPU Space — no GPU at all, at about 5.5–6 tokens/sec. Try it yourself.
※ VIDRAFT-optimized serving. Throughput (tok/s) published after measurement. GLM-5.2: capacity verified, speed under verification. GPQA methods (greedy/maj@8) differ — compare with care. Engine internals are proprietary; results only.
VKAE는 빠르게 · VKUE는 어디서나 · VKIE는 가장 많이. 34.7B 모델 하나로 데이터센터 GPU부터 무료 CPU까지 — 모든 수치 실측, 공개 라이브 데모에서 재현 가능.
스포츠카는 가장 빠르고, 경차는 가장 싸고, 기차는 가장 많이 실어나른다. VKIE는 VKAE의 속도와 VKUE의 절감을 동시에 취해 서빙 능력을 극대화한다.
GPU 가속 · 스포츠카. 단일 스트림 1×B200에서 24 → 220 tok/s. 최고 단일 속도.
GPU 절감 · 경차. 34.7B를 GPU 0으로 ~6–7 tok/s 구동. 최저 비용·최대 접근성.
가속+절감 · 기차. 1×B200 동접(256) 서빙 능력(집계). 최대 처리량·원가효율.
세 엔진은 독립적으로도, 결합해서도 씁니다. 배포 환경에 따라 최적 조합이 달라집니다.
| 🏎️ VKAE | 🚗 VKUE | 🚄 VKIE | |
|---|---|---|---|
| 최적화 차원 | 지연 (속도) | 비용 · 편재 | 처리량 (규모) |
| 핵심 효과 | 최속 단일 생성 | 저사양 구동 | 최대 동시접속 |
| 대표 시나리오 | 실시간 대화 · 코딩 | 엣지 · 온프레미스 | 대규모 API · SaaS |
| 대표 실측 | 506.94 tok/s | 35B @ A10G·CPU | 18,057 tok/s |
기준 모델 Ourbox-35B (총 34.7B / 활성 ~3B 희소 MoE). 모델 변경 없음, 품질 손상 없음.
| Engine | Before | After | Gain | Setup |
|---|---|---|---|---|
| 🏎️ VKAE | 24 | 220 | ≈ 9× | 1×B200 · single stream |
| 🚗 VKUE | 5.4 | 20.0 | ≈ 3.7× | 8GB laptop (RTX 5060) · dense→sparse |
| 🚄 VKIE | 24 | 18,057 | ≈ 750× | 1×B200 · 256 concurrent (aggregate) |
순간 노이즈 스파이크를 배제한 재현 가능한 유효 최고 속도 기준 기록입니다. 외부가 주관한 공개 경쟁에서 검증된 객관 지표입니다.
Ourbox-35B — 같은 가중치, 하드웨어만 바뀐다. 데이터센터 GPU에서 무료 CPU까지 하나의 파일로.
| Hardware | tok/s | Axis |
|---|---|---|
| 1× B200 (datacenter) | 18,057 | VKIE · aggregate (256) |
| 1× A10G (cloud GPU) | 126 | VKUE · single stream |
| 8GB laptop (RTX 5060) | 20.0 | VKUE · 3.7× vs dense 32B |
| CPU-Upgrade (8 vCPU, no GPU) | ~17 | VKUE |
| 🆓 Free CPU Space (2 vCPU, no GPU) | ~6–7 | VKUE · zero cost |
| 1× A100 · 96 동접 | ~460 | VKIE · 단일 대비 3.6× (집계) |
※ 품질은 전 티어 동일 — GPQA Diamond 86.4% (Ourbox-35B, maj@8) ~ 90.9% (Darwin-398B). '빠르고 싸다 ≠ 멍청하다.' 멀티모달 확장: Janus-Pro-1B 이미지 생성 저비용 T4에서 1장 ~28초.
데이터를 클라우드에 못 올리는 공공·국방·의료·금융. 폐쇄망 온프렘 CPU에서 프론티어급 추론.
수억 원대 GPU 클러스터 → 카드 한 장, 나아가 무료 CPU. 진입 비용 두 자릿수 배 절감.
개인·스타트업·중소·공공 누구나. 온디바이스·엣지 수요 급증에 그대로 대응.
양자화 모델(GGUF)·전 벤치 수치·baseline 실행법 공개. 누구나 라이브 데모에서 '적용 전' 수치를 재현할 수 있다.
최적화 서빙의 엔진 내부는 비공개. 결과 속도·하드웨어만 공개 — 결과는 열려 있고, 레시피는 닫혀 있다.
Most scanners flag a pattern and say "maybe a vuln — you check." Chitos verifies inside authorized targets, confirms reproducibility, and chains attack paths into one scenario.
Successor to VIDRAFT Mythos. Where Mythos was static analysis, Chitos verifies.
Python, JS, Go, Java, C/C++, Rust, PHP, YAML — injection sinks, deserialization, credential exposure, crypto misuse, prototype pollution, supply-chain risk. Unproven warnings are filtered out as noise.
Cross-references CVE databases, exploit advisories and public PoCs to explain WHY a finding is dangerous in the current threat landscape. (Requires a Claude API key.)
On authorized targets only — XSS, SQLi, path traversal, command injection. Finds reproducible evidence, retries variant paths when blocked, and links confirmed vulnerabilities into a kill-chain threat map.
Five representative attack scenarios, one click each. Live-test against IBM intentionally-vulnerable demo.testfire.net.
No install. No signup. Your code is never sent to an external API.
Active exploit features may be used only on targets you own or are explicitly authorized to test. Unauthorized testing or scanning may be illegal; all legal responsibility rests with the user.
Simon-algorithm attacks executed on IBM's 156-qubit ibm_kingston — at record hardware scale.
QuantumOS opens along three distinct tracks — console, cryptanalysis, and leaderboard.
Quantum-circuit simulator and console. Build and run algorithms and visualize results directly.
A real-hardware demo analyzing block-cipher structures (Even–Mansour · Feistel) up to N=10 on IBM quantum machines.
Quantum-algorithm benchmark rankings — reproducible results, openly compared.
Drawing a rocket on paper and actually launching it are two different things. Real quantum computers are noisy — a slightly longer computation drowns the answer. For years the public record on real hardware stayed stuck around N≈4.
VIDraft touched that wall. Taking the two most vulnerable structures — Even–Mansour and Feistel — onto IBM's real quantum computer, the team pushed the simplest block-cipher structure to N = 5, 6, 7, 8, 9 and 10, recovering a fresh independent key each time to prove it was not a rigged demo tuned to one answer — using noise mitigation only, no costly error correction.
IBM ibm_kingston (156 qubits) · 15–30 qubit circuits with hundreds of two-qubit gates.
Secret-key recovery on real hardware from N=5 up to N=10 — exceeding the previous public record of N≈4 (to our knowledge).
Hidden-period recovery at block sizes 6 and 8, rank-1 clean.
Error mitigation only — dynamical decoupling, gate/measurement twirling, readout-error correction (no quantum error correction). Independent control keys were tested per instance to confirm the attack isn't tuned to one answer.
Genuine quantum algorithms covering the major cryptographic design paradigms. The two Simon-based constructions were pushed to hardware.
Key recovery in a single query.
Executed on real hardware ✓
Quadratic-speedup key search.
Existential forgery.
Executed on real hardware ✓
美 반도체 권위지 SemiEngineering이 우리 양자 연구를 편집 선정했습니다.
This is a record-scale hardware demonstration — NOT a practical quantum speedup (effective difficulty still tracks the classical birthday bound ~2^(n/2)), NOT a break of AES / RSA / 16-round DES, and NOT quantum-error-corrected (NISQ-era mitigation only). World-first status is unconfirmed pending peer review.
To our knowledge, a first. — A numerical characterization of the tool, not a proof of the Riemann Hypothesis.
VIDRAFT realized, with Blackwell GPUs and AI, the self-adjoint operator that Prof. M. Suzuki (Institute of Science Tokyo) proposed in 2026 as theory only. From its spectrum we derived a closed-form constant (a strong 1σ candidate), a boundary phase, and a prime-onset law — all open and reproducible. The result is Archimedean and universal (shared by all L-functions), not specific to the zeta zeros.
Closed-form fit R²=1.000000 (40 modes) · GUE KS=0.068 vs Poisson 0.349 · von Mangoldt read-back error 0.08% — all reproducible in your browser.
Honest scope: a numerical characterization of the tool, not a proof of the Riemann Hypothesis. The closed form is a 1σ candidate (not confirmed); the result is universal across L-functions. Potential fit for the Journal of Number Theory.
Live services running on the Darwin + AETHER platform.
20 evolutionary models
11 non-invasive adapters
1.5–2× throughput
Single-GPU speed
8 images
Sovereign AI for 3.2M
250-region digital twin
Document & handwriting
Real-time 3D avatar
New-material AI design
Drug toxicity prediction
5-stage autonomous drug pipeline
Quantum algorithm rankings
Quantum sim console
Real HW cryptanalysis
AI code security scan
Fully local AI office
Multi-agent code AI
NL → KS F 1501 drawings
Official collections managed under FINAL-Bench · VIDraft organizations.
Full introduction to the Darwin + AETHER ecosystem (KO/EN). Darwin 1.7B–398B · 11 patents · 9 live services · Blackwell 32-GPU infrastructure · metacognition tech overview.
"The child surpassed both parents — that is evolution." Complete collection of 20 evolutionary language models from Darwin-2B to Darwin-398B. Linked with the arXiv 2605.14386 paper.
11 non-invasive adapters that add self-error detection to major Darwin models without changing the original structure.
"Fastest single-GPU serving of open models via VKAE." Inference-acceleration models achieving 1.5–2× throughput on a single GPU. OpenAI-compatible API, ready to use.
Quantum computing algorithm performance leaderboard. Includes real IBM Heron QPU measurements. NIST randomness verification · Bell inequality violation · VQE · QRAM experiment results.
Achieves 1.5–2× throughput vs. baseline on a single GPU.
The VKAE breakthrough stack delivers 1.5–2× the throughput of standard serving for a 35B-A3B model on one B200 GPU.
Single-GPU serving speed rankings. Darwin · Qwen model throughput comparison. OpenAI-compatible API endpoint provided.
8 Docker images verified for K-AI leaderboard submission. Bake-in method (works without internet on the evaluation server). OpenAI-compatible API.
Public administration AI based on K-AI #1 model — Korean sovereign AI for 3.2M citizens.
Korean sovereign AI for 3.2M citizens. Administration · law · tourism · civil affairs · transport · culture · statistics · agriculture guidance. Analyzes 10+ file formats (HWP·PDF·Excel·images). Includes free citizen AI education. Based on JGOS-31B-Citizen.
52M-person simulation based on KOSIS data. 250 cities/counties × 17 metropolitan regions, 3-tier analysis. Quick (10–30s) / Regional (1–3min) / Deep (5–15min) three stages. 9-Agent hallucination-prevention engine. Generates policy-impact interviews via persona matching.
Polaris Global 14× Champion · From autonomous drug discovery to new material design.
FastAPI + Docker based. Connected to the Darwin-398B-AX production backend. Handwriting recognition · MSDS hazard extraction · health-checkup reading. KIHA SFR-004/001 industrial-safety demo. PDF·PNG·JPG·WEBP support.
MatterGen→MLIP→UQ→4-DB novelty check→Hull→IBM Heron QPU cross-validation 7-gate pipeline. Generative AI + real QPU integration. Compatible with Korea IonQ/KISTI. Paper + patent linkage.
Predicts 5 indicators: Absorption · Distribution · Metabolism · Excretion · Toxicity. Based on the Polaris global-leaderboard 14× champion model. BRICS molecule generation + Boltz-2 docking. tox21-v4 scaffold-holdout AUC 0.790.
RAG→SMILES generation→CythoS validation→SOTA ranking→report. Covers 14 indications. Built-in autonomous retry loop.
IBM Heron QPU measured · NIST randomness 6/6 passed · Real quantum experiment integration.
Genuine quantum-algorithm cryptanalysis suite executed on IBM Kingston real hardware — Even-Mansour N=10 and 3-round Feistel (DES-style) blocks 6/8, rank-1 clean. Honest scope disclosure: not real AES/RSA, not quantum supremacy.
Quantum circuit design · simulation · execution. Visualizes real IBM QPU measurement data. NIST randomness (6/6 pass) · min-entropy 0.974 · Bell inequality violation (2.371) · VQE · QRAM experiments. Run directly in the #console tab.
Quantum algorithm and simulator performance comparison leaderboard. Real QPU measurements shown side-by-side with simulation results. HF Docker metadata auto-updated. 27 likes.
AI security, code agent, local office, legal AI — productivity tools collection.
AI-based code security vulnerability detection. OWASP Top 10 · dependency analysis · secret-key exposure check. Darwin-backbone security fine-tuning.
Browser-based, works without internet. Zero external data transmission. HWP·DOCX·PPT·XLSX·PDF support. MIT license. Released 2026.04.19.
Multi-agent code AI. 1.5M+ MAU verified. Darwin-backbone swarm collaboration. SOMA multi-intelligence architecture.
Case-law-grounded accurate legal AI. RAG + Darwin legal fine-tuning removes legal hallucination.
AI paper writing & enhancement agent. Automates draft generation, reference organization, and improvement suggestions.
22,400+ model sigma.js interactive visualization. Darwin·Trending·Multimodal 5 clusters. Galaxy / Atlas 2 modes.
From a Korean sentence to a verified floor plan — or from a parcel address to a code-compliant 3D building. Drawings, BIM, cost and fire-safety in minutes.
Natural language becomes KS F 1501 drawings, 3D, walkthrough and quantity takeoff — designed by a 397B-parameter MoE and reviewed by a VLM critic.
Natural language becomes KS F 1501 drawings (parallel drafts + AI review) — mm, 1:100, wall poché, area table.
Building code · fire safety · barrier-free access · design metrics · cost estimate · AI review score.
Type a parcel address: zoning, official land price and lot boundaries fetched from the national V-World open-data platform.
Statutory coverage/FAR enforced geometrically; north solar-setback cut into the 3D mass; FAR slider recomputes area, floors and cost live.
Zoning, privacy, south exposure, circulation and proportion scored 0–100 by deterministic rules — not AI self-assessment.
DXF (AutoCAD) · PDF (A3 sheet) · IFC (BIM) · XLSX quantity takeoff.
First-person walk mode, monthly/hourly sun study, and AI photoreal renders and aerials.
※ AI concept-design reference only — coverage/FAR use statutory maximums (actual limits depend on local ordinances), official land price is not market price, and permit and construction documents proceed with a licensed architect.
공개 서비스 실시간 헬스체크 + K-AI 리더보드 Top 10. 4분마다 자동 갱신.
출처: leaderboard.aihub.or.kr · 우리 모델은 ⭐ 강조.
AI essays auto-written daily from news curated by VIDraft.
POCKET runs a 35B model on a GPU-less PC and on a phone with stock llama.cpp — no fork. Measured against Bonsai on the same machine: 2.69× faster generation on CPU, 2.22× on GPU, quality tied. And we published the one row where Bonsai wins.
A plain-language walk through our arXiv paper: Even-Mansour period recovery pushed from N=4 to N=10 on real IBM hardware — and, just as importantly, the six limits the authors wrote down themselves. It does not break AES or RSA, and the paper says so first.
11 models generated and derived from the Darwin Platform are in the K-AI Leaderboard Top 20. This isn't the lucky success of a single model — it's a signal that a system for repeatedly building and deriving models is working.
Inference acceleration extracts more work from the same chip through software. VKAE (VIDRAFT Kernel Acceleration Engine) recorded up to 23.4× higher throughput than the conventional approach on the same GPU. Fast, with quality intact.
Chitos doesn't drop a finding straight into the report. Within authorized scope, it verifies directly, confirms reproducibility, and links attack paths into a kill chain. 3 stages: static scan → threat research → active exploit verification.
http://demo.testfire.net (IBM-authorized vulnerable demo site)
#1 worldwide on 14 benchmarks across 5 fields: efficacy · solubility · toxicity · oncology kinase · ADME.
A sparse MoE keeps only ~3B of 34.7B parameters active, so frontier-class reasoning runs on a single consumer card, a laptop, or a GPU-less CPU server. The VKUE story of lowering the barrier to AI.
AI insights · technology philosophy · industry analysis from VIDraft's CEO
Read more on Brunch @seawolf →Chimera merges and distills open models into stronger ones — verified with no quality loss, and independently confirmed at ICML.
Chimera joins models from different architecture families. That is normally impossible without retraining — tensor shapes and functional roles do not line up. Our filing on compatibility-scored heterogeneous crossbreeding (29 claims) is what makes it a procedure rather than a lucky accident: every tensor pair is scored, then sorted into what can be crossbred and what can only be transplanted.
Filed 2026-05-08 · part of the 12-application portfolio (317 claims)
Chimera = fusing several already-good AIs — without discarding their abilities — into one stronger model. VIDRAFT’s model crossbreeding & merging technology.
Every layer of an LLM is Attention + FFN (its knowledge). Chimera designs the attention itself and transplants another model’s FFN 100% intact — and it supports Mixture-of-Experts.
Conceptual overview only — the internal attention design and gating remain proprietary.
Only measured facts. Internal recipes stay proprietary; the results are public.
Mixing parent models like paint dilutes each strength. Instead we freeze each parent’s core intact and train only a small router that picks the best parent per question.
Result: a 2-expert Chimera (4B) kept GSM8K math at 72.5% — identical to the parent (degradation 0.0). Same-family parents merge to a neutral gain; the real lift comes from cross-family breeding plus the method on the right.
The way to get smarter is to train only on just-right-difficulty problems — dropping the too-easy (already known) and the too-hard (never solved), focusing on the ones solvable with a bit more effort.
Result: Darwin-V9-Chimera-4B lifted Korean KMMLU (6 subjects) from 42.9% to 48.3% (+5.4%p), with no forgetting. The principle — data selection over hyperparameters — was independently confirmed by ICML 2026 work. Measured on a 240-item sample; larger-eval confirmation in progress.
Would combining the best AIs make them stronger? Usually, blending them blurs each one’s strengths. Chimera keeps each model’s strengths fully intact and lets a small “conductor” pick the right expert per problem — so ability never collapses. Add VIDRAFT’s “learn only the right-difficulty problems” method, and skill rises without forgetting (+5.4%p on a Korean reasoning benchmark) — a principle independently backed by ICML 2026 research.
We publish the concepts and the measured results. The internal merge recipes, filter thresholds and gate formulas stay proprietary. Every number here is measured — no inflated claims.
A related research thread — experts cross-checking each other to reduce hallucination — lives under AETHER. Open the AETHER page to read more.
합체로 만든 개체를 자체 정제 파이프라인으로 끌어올린 결과물입니다. 가중치와 수치를 공개해 누구나 재현할 수 있습니다.
Chimera 4B 계보를 자체 정제해 한국어 지식 추론을 끌어올린 모델. KMMLU 6과목 240문항 held-out(greedy)에서 42.9% → 48.3% (+5.4%p).
※ 240문항 기준 표준오차 약 ±1.4%p — 더 큰 셋에서 재확인 필요. KMMLU 6과목 한정 수치이며 전 도메인 일반화를 뜻하지 않습니다. 4B 모델로 플래그십과 직접 비교 대상이 아닙니다. 합체·정제의 내부 설계는 비공개(결과만 공개).
Every officially released Darwin model and its community derivatives — counted by cumulative all-time downloads.
All-time cumulative downloads · HuggingFace shows only the last 30 days by default · updated
Released on VIDraft & FINAL-Bench.
Quantizations, fine-tunes & derivatives built by the community on Darwin.
Notable non-Darwin models released by VIDraft & FINAL-Bench.
Production-ready containers on Docker Hub (vidraft).
Interactive demos on Hugging Face Spaces.
Benchmarks & datasets published by VIDraft & FINAL-Bench.