AI BriefingPyTorchKR news
The latest AI and developer news from PyTorchKR, with short summaries.
Visit website →Red Hat Publishes Performance Comparison Commentary on vLLM vs SGLang
While performance is similar under default settings, the optimal engine varies depending on output length and prefix reuse.
PickStrata Engine Released to Run Qwen3.8-Flash-Next 125B on 12GB VRAM
Generation speed improved 4.6x compared to llama.cpp, enabling 125B MoE model execution with 12GB VRAM
Local Search for Photos, Videos, and Code Using Natural Language
DigUp is a desktop app released under the MIT license that uses EmbeddingGemma 2 to search photos, PDFs, videos, and code locally without the cloud.
VideoJev Releases System One Model for Video Segment Finding and QA
An Apache 2.0 open-source model that finds video segments by scoring candidates without text generation.
Liquid AI Releases LFM2.5-24B-A2B MoE Model for Local Agents
Runs on laptops with 2.45B active parameters, outperforming Qwen3.6 in tool-calling performance
Korea University Releases UltraBERT, Balancing Classification and Retrieval Performance
An English-only model supporting 32K context and licensed under Apache 2.0 for commercial use.
Microsoft Releases StudentSim, a Simulator for Validating Student-Specific AI Tutors
Microsoft Research has released StudentSim, a simulator that mimics individual students' learning patterns, and demonstrated how to use it to validate the instructional quality of AI tutors.
Sakana AI Releases MASS Paper
MASS based on Qwen3.6-27B improves research benchmark performance by 1.2–1.6x, while SW benchmarks decline
Ai2 Cuts Wait Times by 96% with New GPU Scheduling Policy
Introducing GPU time budgets and fair-share allocation instead of priorities significantly improves wait times while maintaining 98% cluster utilization
RSI-Anything: Coupled Policy-World Model Updates via World Models
Research demonstrating applications across various domains, including a 31.2% code repair success rate, notes that the 99% GUI success rate is based on policy ranking.
VKORE-46M-NPU Claims Backpropagation Training on Inference-Only NPU: A Model Card Analysis
Analysis of the VKORE-46M-NPU model card released on Hugging Face, which claims backpropagation training on an inference-only NPU and compares loss trajectories with GPUs.
NVIDIA Releases ReplaySSM to Minimize SSM State Writes
Improves SSM kernel performance by up to 1.8x at batch size 256 and speculative decoding throughput by up to 2.1x at batch size 512
a16z Releases 2026 AI Market Status Report
Hyperscaler capex projected at $777 billion, with agent token usage surpassing human usage.
NVIDIA to Host AI Day 2026 in Seoul
Held at COEX on November 9-10, the event offers RTX Spark experiences and free opportunities to take technical certification exams.
Anthropic Reveals Four Types of Unintended Claude Behaviors
Anthropic disclosed four cases where Claude accessed live websites during evaluations and implemented measures to block live internet access.
Unsloth Releases Guide for Fine-Tuning and Local Serving of LLM-Based Decision Models
Presents a method for training and serving LLMs like Qwen3.5 by attaching decision heads to perform classification and judgment tasks
Microsoft Releases 'Microsoft-Decision-1', a Jev-Compatible Decision Model
Offers a lightweight decision model specialized for classification and routing tasks at $0.042 per 1M input tokens
OECD Releases Report on Agentic AI Adoption Status
Based on interviews with 25 organizations, including 9 in Japan, the report emphasizes phased oversight, requiring human approval for high-risk tasks.
vLLM Releases GuideLLM v0.8.0
Added latency SLO-based max concurrency search and agent trace replay in WEKA/OTEL formats.
REA Releases MCP Integrating Reverse Engineering Tools
An open-source project bundling tools like Ghidra to enable coding agents to analyze executables and the web.