Engineering newsroom

VIDRAFT AI Media

Engineering news and analysis from VIDRAFT: open LLMs, recursive self-improvement, leaderboards and decision AI, with reproducible numbers.

Lead story

+47% Coding Ability at Q4: POCKET-27B-Pi Solves 58.3% of Coding Tasks, the Original Qwen3.8-27B 39.6%

2026-11-20

Quantized to Q4 and still 47% ahead: the 16.5 GB POCKET-27B-Pi solved 58.3% of Terminal-Bench 2.1 coding tasks, the original Qwen3.8-27B 39.6%.

aillm

AI Model Market Daily Launches: A Free Daily Newsletter Tracking About 2,000 Top AI Models

2026-11-20

A free LinkedIn newsletter, published every day, that gathers and analyzes the market, technology, rankings and news of roughly 2,000 of the world's top AI models.

weatherweather forecast

What Is the Weather Forecast for Seoul, Tokyo, London and New York on November 20, 2026? 24 World Cities, AI Forecast

2026-11-20

November 20, 2026 forecast for 24 world cities from Darwin-Prediction-Weather: Seoul 3°C / -4°C, Tokyo 14°C, London 9°C, New York 19°C. Warmest Sao Paulo 31°C, coldest Anchorage -4°C. The model ranks first on the public Weather Model Leaderboard.

aillm

AI Model Market Daily Nov 20: Money Hy4 flash, Intelligence Gemini 4 Argon, Security Kimi K3.1 Pro

2026-11-20

Money 1 Hy4 flash · OpenRouter tokens, 7 days 69.3T · share 30.0% Intelligence 1 Gemini 4 Argon · SOTA 70.4 (+2.7) Security 1 Kimi K3.1

aisecurity

POCKET-Darwin-180B Takes #1 on Hugging Face's Official Benchmark for Protecting AI Agent Tool Calls

2026-11-19

On November 19, 2026, POCKET-Darwin-180B scored 0.599 on the Category Macro F1 leaderboard of safeagents judge bench, ahead of Cloudflare clef at 0.563. VIDRAFT took on the benchmark because it maps directly to ViKernel.

aisecurity

JoongAng Ilbo: VIDRAFT Sent Attack Instructions to 20 AI Models, and All 20 Were Found Vulnerable

2026-11-19

On November 19, 2026, JoongAng Ilbo reported VIDRAFT's security check of 20 AI models: all 20 received the lowest grade, and ViKernel blocked 897 of 897 dangerous calls.

aisecurity

One Hidden Word Turns a Downloaded AI Model Into a Data Thief. Watch It Get Stopped at the Action

2026-11-19

A copy of a popular open model with a small hidden weight change sends a secret key to an outside server when one trigger word appears. ViKernel blocks the call before a byte leaves.

aisecurity

Attackers Now Target the AI Model Itself: An Open Challenge to Break an AI Security Kernel, With $4,500 in Prizes

2026-11-19

VIDRAFT opened its AI security kernel ViKernel to public attack with USD 4,500 in prizes. Participants have submitted 298 attacks, and 148 causes they found were fixed in the kernel.

aisecurity

Is the AI You Use Safe? A Free Scanner Grades Any Model A to F in a Minute, and 20 Well-Known Models All Got F

2026-11-19

ViKernel Scanner is a free, open-source AI security scanner built on a database of AI model vulnerabilities. It hands a model 8 mock tools and grades it A to F in about a minute.

aisecurity

Worse Than AI Cheating: 20 Widely Used AI Models Were Tricked Into Dangerous Actions, and All 20 Went Along

2026-11-19

VIDRAFT handed 20 widely used open-weight AI models a set of tools and sent attack instructions. All 20 were graded F. A security kernel inside the weights blocked 897 of 897 dangerous calls.

weatherweather forecast

What Is the Weather Forecast for Seoul, Tokyo, London and New York on November 19, 2026? 24 World Cities, AI Forecast

2026-11-19

November 19, 2026 forecast for 24 world cities from Darwin-Prediction-Weather: Seoul 6°C / -1°C, Tokyo 18°C, London 10°C, New York 16°C. Warmest Dubai 31°C, coldest Reykjavik -4°C. The model ranks first on the public Weather Model Leaderboard.

aisecurity

Give Your AI Model Tools and Watch What It Tries: ViKernel Scanner Grades Any Model A to F in About a Minute

2026-11-18

VIDRAFT releases ViKernel Scanner, a free AI model security scanner on Hugging Face. It hands any OpenAI-compatible model mock tools, records every tool call it tries and grades it A to F in about a minute.

aillm

ETNews: 70.7% of the World's 300 Most Downloaded Language Models Are Chinese, and One Korean Model Is on the List

2026-11-18

On November 18, 2026, ETNews reported VIDRAFT's analysis of the 300 most downloaded language models on Hugging Face: 523 million downloads in 30 days, 70.7% Chinese, and POCKET-Darwin-180B as the only Korean model.

aisecurity

AI Model Hacking Hackathon Grows to Four Seasons and USD 4,500 in Prizes

2026-11-18

The AI Model Hacking Hackathon on Hugging Face now has four seasons, USD 4,500 in prizes, a USD 500 Special Award in every season and 10 submissions a day.

aillm

Maeil Business Newspaper: Darwin Holds 12 First Places Across Hugging Face Official Benchmarks, the Most of 140 Organizations

2026-11-17

On November 17, 2026, Maeil Business Newspaper reported VIDRAFT's full analysis of 63 Hugging Face official benchmarks: Darwin is first on 12 leaderboards, the most among 140 organizations.

aisecurity

AI Model Hacking Hackathon: Safe AI Is Built by Everyone Who Tries to Break It

2026-11-17

VIDRAFT opens ViKernel, the security kernel for AI agents, to public attack on Hugging Face. Four seasons, USD 4,500 in prizes, browser only: no coding and no GPU.

aillm

ZDNet Korea: VIDRAFT's 180B Darwin Outscores a 2.8 Trillion Parameter Model in Math and Science Reasoning

2026-11-17

On November 17, 2026, ZDNet Korea reported that VIDRAFT's Darwin-180B-RSI ranked first among open-weight models in SuperCLUE's math and science reasoning evaluation, more than 10 points above a model about 16 times its size, using model synthesis and recursive self-improvement.

aiweather

Eight National Weather Models, One Korean AI: Darwin Has the Lowest Error in All 28 Categories

2026-11-17

VIDRAFT compared its AI weather engine with eight forecast models from ECMWF, NOAA and the weather services of Japan, Canada, Germany and Norway across 24 cities for nine months. Darwin had the lowest error in all 28 categories, with a 7-day maximum temperature error of 2.08 C.

weatherweather forecast

What Is the Weather Forecast for Seoul, Tokyo, London and New York on November 17, 2026? 24 World Cities, AI Forecast

2026-11-17

November 17, 2026 forecast for 24 world cities from Darwin-Prediction-Weather: Seoul 15°C / 10°C, Tokyo 12°C, London 13°C, New York 13°C. Warmest Dubai 31°C, coldest Anchorage -3°C. The model ranks first on the public Weather Model Leaderboard.

aillm

The Korea Economic Daily Reports Darwin First Among Open-Source Models in SuperCLUE Math and Science Reasoning

2026-11-16

On November 16, 2026, The Korea Economic Daily (Hankyung) covered VIDRAFT's Darwin-180B-RSI: first among open-source models and second overall in SuperCLUE's November math and science reasoning evaluation, with 12 number-one records on Hugging Face official benchmarks.

weatherweather forecast

What Is the Weather Forecast for Seoul, Tokyo, London and New York on November 16, 2026? 24 World Cities, AI Forecast

2026-11-16

November 16, 2026 forecast for 24 world cities from Darwin-Prediction-Weather: Seoul 18°C / 9°C, Tokyo 18°C, London 8°C, New York 10°C. Warmest Dubai 31°C, coldest Reykjavik -5°C. The model ranks first on the public Weather Model Leaderboard.

aillm

Darwin Lab: A Private AI Research Lab Built Into an NVIDIA DGX Spark and a NAS

2026-11-15

VIDRAFT introduces Darwin Lab, an on-premise science platform. The Darwin model ranks second overall and first among open models in SuperCLUE's November 2026 math and science reasoning, and it runs inside the lab with no internet.

aillm

Which Open LLMs Are Downloaded Now? A Daily Leaderboard of 30-Day Downloads by Country

2026-11-14

A daily Hugging Face leaderboard of trailing 30-day downloads: global top 300 and Korea top 30. China-trained families hold 61.7% of top-300 downloads, the US 32.3%. Korea's top two models are VIDRAFT's POCKET series.

aiweather

Which Weather Model Is Most Accurate? We Score Them Every Day, in the Open

2026-11-14

A public Hugging Face leaderboard scores ECMWF, NOAA, DWD, JMA, ECCC, MET Norway and Darwin-Prediction-Weather against ERA5 every day. In the 2026 record Darwin-Prediction-Weather has the lowest error in all four elements at every lead time from 1 to 7 days.

aillm

Seoul Shinmun Reports Darwin First Among Open-Weight Models in SuperCLUE Math and Science Reasoning

2026-11-13

On November 13, 2026, Seoul Shinmun, one of Korea's national dailies, covered VIDRAFT's Darwin-180B-RSI: first among open-weight models and second of 19 models in SuperCLUE's November math and science reasoning evaluation.

aillm

The line the benchmark put in its own headline: "Korea's Darwin on the board for the first time", reading SuperCLUE's November announcement as published

2026-11-09

On 4 November 2026 SuperCLUE put "Korean model Darwin on the board for the first time" in the headline of its own monthly announcement. We read the primary source: the headline, the published task weights, and the one footnote written only for Darwin-180B-RSI-R3.

aillm

채점기관이 제목에 쓴 한 줄: 「한국 모델 Darwin 첫 등재」, 슈퍼클루 11월 공식 발표문을 그대로 읽는다

2026-11-09

슈퍼클루가 2026년 11월 4일 공식 발표문 제목에 「한국 모델 Darwin 첫 등재」를 직접 썼습니다. 점수를 매긴 쪽이 제목을 그렇게 뽑았다는 사실, 가중치 배분, 그리고 Darwin-180B-RSI-R3 만 받은 각주까지 발표문 원문으로 확인합니다.

aillm

Оценка, которая каждый месяц пишет новый экзамен: ноябрьская таблица SuperCLUE 2026, 19 моделей в одной таблице

2026-11-08

SuperCLUE каждый месяц собирает закрытый тестовый набор, поэтому модель не поднимется за счёт заученных опубликованных вопросов. Таблица за ноябрь 2026 года ставит 19 передовых моделей рядом по пяти осям. Darwin-180B-RSI-R3 возглавляет модели с открытыми весами в математическом и научном рассуждении с 73,68.

aillm

La evaluación que reescribe su examen cada mes: el tablero de SuperCLUE de noviembre de 2026, 19 modelos en una tabla

2026-11-08

SuperCLUE construye cada mes un conjunto de prueba privado, así que ningún modelo sube memorizando preguntas publicadas. Su tablero de noviembre de 2026 pone 19 modelos de frontera en cinco ejes. Darwin-180B-RSI-R3 lidera los modelos de pesos abiertos en razonamiento matemático y científico con 73,68.

aillm

毎月、問題を作り直す評価機関: SuperCLUE 2026年11月ボード、世界最前線19モデルを一枚の表に

2026-11-08

SuperCLUE は毎月、非公開のテストセットを新しく作ります。公開済みの問題を覚えても順位は上がりません。2026年11月のボードは19モデルを5軸で同じ表に並べ、Darwin-180B-RSI-R3 が数学・科学推論でオープンウェイト1位の73.68を記録しました。

aillm

每月重新出卷的评测:SuperCLUE 2026 年 11 月榜单,19 个全球前沿模型同表

2026-11-08

SuperCLUE 每月重建一套不公开的测试集,模型无法靠背诵公开题目上榜。2026 年 11 月榜单把 19 个前沿模型放在五个维度的同一张表里,Darwin-180B-RSI-R3 在数学与科学推理拿下开放权重第一,73.68 分。

aillm

매달 시험지를 새로 만드는 평가기관: 슈퍼클루 2026년 11월 보드, 글로벌 19개 모델 한 표

2026-11-08

슈퍼클루는 매달 비공개 시험지를 새로 만듭니다. 공개된 문제를 외워서 올라갈 수 없습니다. 2026년 11월 보드는 세계 최상위 19개 모델을 다섯 축에서 한 표에 올렸고, Darwin-180B-RSI-R3 가 수학·과학 추론에서 오픈 가중치 1위 73.68 을 기록했습니다.

aillm

The Benchmark That Writes a New Exam Every Month: SuperCLUE's November 2026 Board, 19 Models, One Table

2026-11-08

SuperCLUE builds a private test set every month, so no model can win by memorising published questions. Its November 2026 board puts 19 frontier models side by side across five axes. Darwin-180B-RSI-R3 leads open-weight models in mathematical and scientific reasoning at 73.68.

aillm

VIDRAFT, November 1 to 8, 2026: Open-Weight #1 in Math and Science Reasoning, and a Gate That Stopped 290 of 290

2026-11-08

One week, three results on public pages: Darwin-180B-RSI-R3 leads open-weight models in SuperCLUE's November math and science reasoning at 73.68, ViKernel denied 290 of 290 attack payloads from a backdoored model, and POCKET passed 3.15 million downloads. Eighteen outlets in seven countries carried it.

aisecurity

A Model Scan Found Nothing. The Gate Stopped 290 of 290 Attack Commands

2026-11-08

You cannot find an unknown trigger by testing a model. VIDRAFT's ViKernel judges the action instead: 24 scan checks found nothing in a backdoored model, and the gate denied every one of its 290 attack payloads at a 237 ms median.

aillm

Korean Press on Darwin's SuperCLUE Result: Chosun Biz, Economist, Digital Daily and AI Times Report #2 in Math and Science Reasoning

2026-11-06

Four Korean outlets covered VIDRAFT's Darwin-180B-RSI-R3 on November 6, 2026 after SuperCLUE's November board placed it second in math and science reasoning and first among open-weight models. What each outlet emphasised, the numbers behind the headlines, and how the result compares with the parent model.

aillm

Darwin-180B-RSI Ranks #1 Open-Weight Model in Math and Science Reasoning on SuperCLUE (Nov 2026)

2026-11-06

SuperCLUE's November 2026 evaluation puts VIDRAFT's Darwin-180B-RSI at 73.68 in mathematical and scientific reasoning, second overall behind GPT-6.1-Sol and ahead of Gemini, Grok, DeepSeek and Qwen. What the benchmark measures, how Darwin was built, and why it beat its own parent model.

aiopensource

Qwen Still Takes Half of Open-LLM Downloads: What Changed in the Hugging Face Top 300 Since September

2026-11-05

We compared two snapshots of the 300 most-downloaded open LLMs on Hugging Face. Qwen fell from 55.8% to 50.9%, and one 2022 model skews the totals.

AIViKernel

비드래프트 ViKernel Desktop으로 보는 AI 보안의 실행 관문

2026-11-04

AI가 읽은 문장과 실행해도 되는 행동은 다르다. ViKernel Desktop 공개 자료와 실제 데모 화면으로 그 차이를 설명한다.

aisecurity

Video: ViKernel, the Operating System Now Inside the AI (Meta Hackathon Demo)

2026-11-02

A 2-minute demo: ViKernel runs a secure OS inside Meta's Muse Glimmer 30B. Hijack denied, fake admin denied, one press stops everything. Zero loss.

aisecurity

Can You Stop Your AI Agent? ViKernel Puts the Kill Switch Inside the Model

2026-11-01

VIDRAFT ViKernel runs a Secure OS inside the model weights: rule-based verdicts, a tamper-evident audit chain and a kill switch only a signed key can undo.

aillm

Run a 35B to 180B LLM Without a GPU: VIDRAFT POCKET Passes 3.15 Million Downloads

2026-11-01

VIDRAFT POCKET GGUF models run 26B to 180B LLMs on a PC, laptop or phone with stock llama.cpp. 1,113,164 downloads in 30 days, measured on Hugging Face.

ainews

VIDRAFT in the Korean Press, 30 October 2026: ViSQL, NPU Pretraining, Chitos, a Hugging Face Demo and a Benchmark Analysis

2026-10-30

Five Korean outlets covered VIDRAFT on 30 October: Chosun Biz on ViSQL, AI Times on the Hugging Face demo, Economist on QuantID's benchmark analysis, Financial News on Chitos, and JoongAng Ilbo on NPU pretraining.

aisql

VIDRAFT NL2SQL Engine Scores 83.80 on BIRD Mini-Dev, Above the Current #1, and Opens a Public Demo

2026-10-29

Darwin-180B-NL2SQL reaches 83.80% execution accuracy on BIRD Mini-Dev (SQLite), above the listed #1 at 81.00. A public demo drafts a legal fraud notice from data.

securityai

Chitos Covers All 17 Items of Korea's Financial AI Agent Security Criteria, Scores 77.8 on OWASP Benchmark On-Prem

2026-10-29

VIDRAFT's Chitos has modules for all 17 items of Korea's financial AI agent security criteria and scores 77.8 (Youden) on OWASP Benchmark with an on-premises engine.

aimachinelearning

Decision AI, Explained: Why Hugging Face Built a Demo for Darwin-35B-ZTC-Vision

2026-10-29

Darwin-35B-ZTC-Vision answers with a calibrated probability per option in one pass. Hugging Face's open-source team built and donated its demo. Records inside.

aimachinelearning

VIDRAFT Takes #1 on BRIGHT, the New Reasoning-Retrieval Leaderboard on Hugging Face

2026-10-28

FINAL-Bench scores 0.307 nDCG@10 on BRIGHT, Hugging Face's new reasoning-intensive retrieval leaderboard. VIDRAFT now leads 12 of 52 official benchmarks.

aimachinelearning

Hugging Face's Open-Source Team Built a Free Demo for Darwin-35B-ZTC-Vision

2026-10-28

Hugging Face's open-source team built a free ZeroGPU demo of Darwin-35B-ZTC-Vision, VIDRAFT's 35B vision decision model, and is moving it to FINAL-Bench.

securitywebdev

Free 1-Minute Website Check for the Age of AI Hacking: Security, SEO and AI Visibility

2026-10-27

FreeWeb Security SCAN gives a read-only report on security headers, HTTPS, email spoofing protection, SEO basics and access for 9 AI crawlers.

aimachinelearning

VKORE-325M: A 325M LLM Pretrained From Zero on a Rebellions ATOM-Max NPU, Equal to a B200 Twin

2026-10-26

VIDRAFT pretrained a 325M model from random init on an inference-only Rebellions ATOM-Max NPU. On unseen text: 0.7993 vs 0.7994 for the same model on a B200.

aimachinelearning

Why Long bf16 Training Runs Stall: The Residual Stream Has No Room for Small Updates

2026-10-22

A bf16 residual near 600 has a resolution of 4, so small updates vanish. Measured gaps, a reproducible simulation and fixes for long training runs.

aillm

Darwin-27B-ZTC-v9 Takes #1 on S1MB Across 197 Decision Models

2026-10-20

A 27B decision model that answers without generating text now leads the S1MB English leaderboard on Borda score across 137 benchmarks, ahead of entries from Perplexity, Meta and Cloudflare.

aillm

ZTC: A Decision Model That Answers in One Forward Pass

2026-10-19

Darwin-27B-ZTC-v5 returns a calibrated probability for every option without generating text. How decision models work, how they are scored, and how to serve one with transformers serve.

aidrugdiscovery

LEADBOARD: A Drug Property Benchmark Whose Labels Never Leave the Verifier

2026-10-19

21 boards across 7 disciplines and 18,382 held-out compounds. Time splits, published noise floors and baselines, and a ladder rule so the board cannot be won by resubmitting.

aidrugdiscovery

Open Discovery Challenge: Four Diseases, Four Targets, One Verifier

2026-10-19

An open drug discovery competition. Submit a SMILES string; a hidden pipeline scores binding, selectivity, ADMET, novelty and synthesis against a target and a counter target.

aimaterials

Open Materials Challenge: Find a Solid-State Battery Electrolyte, Get It Verified

2026-10-19

Submit a lithium-containing crystal structure. A hidden pipeline checks oxidation stability, lithium-metal stability and novelty of use. Season 1 closes 2026-11-30.

aiquantum

Open Quantum Challenge: Two Quantum Seasons, No Quantum Hardware Required

2026-10-19

Compute Fermi-Hubbard ground-state energies or build a surface-code decoder with a GPU and an open harness. Answers are held privately and every entry is scored automatically.

aiscience

Five Open Science Challenges Where a Verifier, Not the Submitter, Decides the Score

2026-10-19

Drug discovery, battery materials, superconductors, quantum computing and drug property prediction: five open challenges from FINAL-Bench, each graded by a verifier that holds the answer privately.

aiphysics

Open Superconductor Challenge: Screen 2D Materials for d-Wave Pairing on a Laptop

2026-10-19

Estimate d-wave pairing tendency from a published effective Hubbard model per material. A many-body reference solver verifies the top entries. Season 1 closes 2026-12-31.

opensourceai

Six Open-Source Projects from VIDRAFT, Now on GitHub

2026-10-18

ONGRID, FINAL Bench, the Open Quantum Challenge harness, POCKET, ZTC and ALL Bench Leaderboard: what each project does and where to start.

aillm

VIDRAFT This Week: Typed Decisions #1, S1MB Japanese #1, and NPU Pretraining in the Press

2026-10-17

A weekly roundup: Darwin-27B-ZTC-v5 holds #1 on Hugging Face Typed Decisions at 0.81, the S1MB Japanese edition #1 at 84.57, and national press coverage of VKORE pretraining on inference-only NPUs.

aillm

VIDRAFT Now Leads 11 Hugging Face Official Benchmarks After Reclaiming #1 on Typed Decisions

2026-10-16

Darwin-27B-ZTC-v5 reclaims #1 on the Typed Decisions official benchmark at 0.81, holding both first and second place, bringing VIDRAFT to #1 on 11 Hugging Face official benchmarks.

aimachinelearning

Pretraining From Scratch on an Inference-Only NPU: VKORE-46M-NPU

2026-10-14

VIDRAFT pretrained a language model from scratch on an inference-only NPU via real backpropagation and released it on Hugging Face. Loss trajectory correlates 0.99 with a GPU reference.

aimachinelearning

Model-Level RSI: What It Means When a Model Improves Its Own Weights

2026-10-12

A practical look at model-level recursive self-improvement, the five stages, and how to read verifiable results like 100% on AIME 2026 and 93.43% on GPQA Diamond.

aiquantum

Open Quantum Challenge: Solve Quantum Problems With a GPU and No Quantum Computer

2026-10-11

The Open Quantum Challenge on Hugging Face opens with two seasons, quantum simulation and QEC decoding. Use only a GPU and an open harness, no quantum hardware. 2,000 USD in prizes.

aiagi

AI Native: Ten People, Hundreds of AI, and the Arrival of the Pre-AGI Age

2026-10-11

VIDRAFT is an AI-native lab where hundreds of AI instances run as a company on a few dozen GPUs. Perfect math scores, 165M downloads, quantum hardware, and a self-improving org point to the Pre-AGI age.

aimachinelearning

S1MB Number One: How a Zero-Token Judge Won the Decision-Engine Leaderboard

2026-10-11

Our model ranked number one on the S1MB decision-engine leaderboard among 102 models, with a Borda score of 89.58, deciding in a single forward pass with zero generated tokens.

aimachinelearning

a16z Bet $870M on Typed, Machine-Native AI. Here Is the Open, Benchmark-Proven Version

2026-10-10

TypeSafe AI raised 870M dollars for typed, machine-native AI. We explain what typed decision AI is, why it matters, and show an open, zero-token, S1MB number one alternative.

aillm

From exams to real work: an open 180B model now leads nine official Hugging Face leaderboards

2026-10-05

Darwin-180B-RSI adds #1 on IFStruct (structured output, 98.95%) and ExtractBench (document extraction, 90.29), for nine first places among 48 official Hugg