AI BriefingPyTorchKR news
The latest AI and developer news from PyTorchKR, with short summaries.
Visit website →NVIDIA Cosmos 3: Four Robot and Autonomous Driving Simulation Models Released via Causal Distillation (Some Achieve Real-Time Performance)
NVIDIA has released four interactive world simulation models based on Cosmos 3 under the OpenMDW-1.1 license.
OpenAI GPT-6 Cyber, Launching Exclusively for Approved Security Organizations
It recorded a 0.07% out-of-scope attack rate while discovering 948 Linux kernel privilege escalation vulnerabilities.
Thinking Machines Lab Releases Active 17B Inkling 2 Small Tinker and Plans to Open Weights
Inkling 2 Small achieved the final score of the first generation in just 75 steps of RL fine-tuning.
CopilotKit Unveils OpenDots, Featuring Dedicated Computers and Document Editing for Each Agent
CopilotKit has released OpenDots in alpha, providing isolated computers per agent and an editable document repository.
Mago Studio Adds Local ComfyUI Rendering Support and H3 Model License Restrictions
Mago Studio is a local-first AI video editor that supports local rendering via ComfyUI, but its default model H3 is restricted for use in South Korea due to licensing terms.
SuperClue November SOTA Evaluation: GPT-6.1-Sol Ranks First in Coding, Kimi-K3.1-Pro Officially Tops Overall
GPT-6.1-Sol ranked first in coding and mathematics, while Kimi-K3.1-Pro took the official overall first place.
Modular LLM Inference Handbook, 57 pages released for free
Modular has released a free 57-page handbook covering LLM inference metrics, optimization, and operational costs.
MIT CSAIL Uses GPT-6 Astra to Directly Edit Weights, Achieving 48.1% on Qwen3.5-0.8B Math Benchmark
Without an optimizer, an LLM directly modified weights to raise Qwen3.5-0.8B's MATH-500 score from 7.9% to 48.1%.
Anatomy: Explaining Concepts with Moving Machines Instead of Boxes and Arrows
Anatomy visualizes concepts as moving machines for use in technical documentation and lecture materials.
Hugging Face Publishes Draft Standard for Transformers Model Integration Contracts
Hugging Face proposed a draft standard for the Transformers Model Integration Contract (MIC) to strengthen CI regression detection.
Google DeepMind Unveils Antigravity Cheminformatics Skills
They unveiled a demo that performs drug design iterations, which typically take days, in about 1 hour.
Nemotron 3.5 Lightning Generation Speed Improved by 1.7x with DSpark on DGX Spark
NVIDIA announced that Nemotron 3.5 Lightning generation speed on DGX Spark is up to 1.73 times faster when DSpark is applied.
109M JEPA Model Demonstrates Superior Alignment with Vision Models Compared to 72B LLM
Theoretical proofs and experiments confirm that a 109M JEPA model aligns better with vision models than a 72B LLM.
OpenAI LifeVal Benchmark: GPT-6 Astra Records 80.6%
OpenAI released the LifeVal benchmark, which measures the ability to perform everyday tasks.
NVIDIA Auto Ontology: Validating Text-to-SQL with DB Metadata-Based Ontologies
NVIDIA has released an open-source framework that automatically generates ontologies from DB metadata to validate AI agent SQL.
NVIDIA Accelerates MLIP Molecular Dynamics by Up to 1.7x Without Retraining
NVIDIA has disclosed an FP32 technique that accelerates MLIP molecular dynamics by up to 1.7x without retraining.
GPT-6 Astra Improves Qwen3.5-0.8B Math Performance 6x via Weight Space Programming
GPT-6 Astra improved the MATH-500 score of Qwen3.5-0.8B to 48.1% by modifying its weights without training.
Mistral Large 4 Achieves Coding Performance Similar to GLM-5.3 with 1T MoE Model
Mistral released the weights for its 1T MoE model, Mistral Large 4, demonstrating token efficiency.
Google Sleep, Qwen3-8B Achieves 79.2% on AIME-24
Sleep improved the AIME-24 score by 5.4%p through offline integration.
ViKernel v0.4 blocks 290 backdoor attacks via GPU circuits within model weights
ViKernel v0.4 transplanted CPU, NPU, GPU, RAM, and ROM circuits into the Meta Muse Glimmer 30B model weights as a sidecar layer, effectively blocking agent security threats.