AI BriefingPyTorchKR news
The latest AI and developer news from PyTorchKR, with short summaries.
Visit website →OpenArm Releases Open-Source Dual-Arm Robot and Evaluation Platform
OpenArm has released an open-source dual-arm robot hardware stack featuring 7 degrees of freedom, along with tools for data learning and policy evaluation.
PyTorchKR Releases Collection of Key AI Papers Including Agent Harness Optimization
Introduced seven latest AI research results, including improving task success rates from 23.3% to 44.3% through agent harness distillation.
Inco AI Releases Splash, an Inference Engine Dedicated to Apple Silicon
Inco AI has released Splash, an Apple Silicon inference engine that applies model-specific kernels and draft models.
UC Berkeley Publishes 'Harness Tax' Study: Up to 5x Cost Difference Between Harnesses
UC Berkeley researchers identified the 'Harness Tax' phenomenon, where achieving the same quality with coding agent harnesses can cost up to 5 times more depending on the harness design.
PrismML Releases 'Ternary Bonsai 2 27B', a Ternary Quantized Model Maintaining 98.2% of FP16 Performance
PrismML has released 'Ternary Bonsai 2 27B', a 27B multimodal LLM quantized to ternary values that maintains 98.2% of FP16 performance.
Transformer Performance: Mechanism Diversity Is Key Over Placement
New research demonstrates that combining diverse mechanisms (especially including SSMs) has a greater impact on performance than the placement order of attention mechanisms in Transformer stacks.
RLT: Proposed Recurrent Decoder Structure with Computational Depth Increasing with Token Length
The RLT structure, where the recurrent path lengthens and computational load increases during token generation, along with preliminary experimental results, has been released.
Recipe Released for Serving 320B GLM-5.3-Flash on Free Kaggle TPUs
An open-source repository has been released for serving the 320B MoE model GLM-5.3-Flash on Kaggle's free TPU v5e-8.
BrowserSkill Released: Agent CLI That Reuses Logged-In Browser Sessions
BrowserSkill, a CLI tool for agents that reuses actual logged-in browser sessions while operating in a separate window, has been released.
PickByteDance Open-Sources 'Bernini', an Integrated Framework for Video Generation and Editing
ByteDance has released Bernini, an open-source framework that combines an MLLM planner and a DiT renderer to perform six video generation and editing tasks.
SenseNova-Vision Released as Unified Model for Multiple Vision Tasks
SenseNova-Vision has been released, integrating detection, segmentation, and depth estimation into a single generative model without task-specific heads.
NVIDIA Accelerates MoE Training 10x with TE for JAX
NVIDIA achieved 1,025 TFLOPs per GPU on GB200 for Dropless MoE training via Transformer Engine for JAX, improving throughput by approximately 10x.
llmfit Released: Pre-estimate Local LLM Speed
The Rust-based tool llmfit has released a feature that estimates generation speed and compatibility from hardware specs before downloading local LLM models.
Unitree Releases UnifoLM-WLA-1.0, a 6B Model for Humanoids
Unitree has released UnifoLM-WLA-1.0, a 6B parameter model for humanoid manipulation.
Paperthin Releases 28 Skills for Coding Agents
Paperthin, a collection of 28 open-source skills designed to resolve discrepancies between documentation and code generated by coding agents, has been released.
Pick9 AI Agent Skills for Three.js Game Development Released
Nine Three.js game development skills usable in Codex and Claude Code have been released, providing self-verification capabilities based on a visual quality scorecard.
PickLocal Control App for ChatGPT 'Chat On Steroids' Released
The open-source MCP app 'Chat On Steroids' has been released, enabling local file modifications and terminal command execution via ChatGPT conversations.
Edge0 Runs 35B MoE Model on 3GiB Memory via SSD Streaming
Edge0 released an inference framework that runs 35B MoE models on 3GiB of memory using SSD expert streaming technology.
A2A Releases Official CLI Tool
The official CLI for the A2A protocol has been released, enabling agent invocation, server execution, and management directly from the terminal.
PickAnalysis of the 'Home Advantage' Phenomenon in Document Extraction Benchmarks
A 'home advantage' pattern was discovered where, among 13 document data extraction benchmarks, the manufacturer's models ranked in the top 3 in all 10 benchmarks they participated in, and took first place in 7 of them.