AI BriefingPyTorchKR news
The latest AI and developer news from PyTorchKR, with short summaries.
Visit website →Nanbeige4.3-33B ranks first on 17 of 20 benchmarks
Kanzhun released the 33B dense model Nanbeige4.3-33B under the Apache 2.0 license.
Sebastian Raschka, Qwen3 0.6B GRPO Experiment: Collapse to 0.4% at 100 Steps with KL Term, Achieves 47.4% Without It
Confirmed that removing the KL term is essential for preventing performance collapse and improving efficiency when training Qwen3 0.6B with GRPO.
Vercel durable-harness supports log-based turn resumption on function failure
Vercel has released durable-harness, which allows coding agent turns to continue via logs.
Microsoft and ThinkingBox-Training Released: Qwen3.8-27B Repeat Success Rate Improved 2.8x
The number of tasks where Qwen3.8-27B achieved 20 consecutive successes increased from 38 to 107.
Memento Achieves 87.88% on GAIA via Case-Based Reasoning Without LLM Fine-Tuning
With LLM weights frozen, it achieved 87.88% on the GAIA validation set by leveraging a case bank.
StepFun Step Code adds built-in sub-agents and web deployment features
StepFun released Step Code, derived from Pi, with built-in sub-agent and web deployment capabilities.
Jane Street and NVIDIA Release Mamba-2 cuTile Kernels, Achieving Up to 1.82x Speedup on B300
cuTile kernels that boost Mamba-2 forward propagation speed by up to 1.82x on B300 GPUs have been released.
Corgi provides coding agent state tracking and session persistence
Corgi improves terminal management efficiency through agent state tracking and background session persistence.
StepFun Step 5 Preview Weights Released, Runnable on 8 H200s
StepFun released the weights for its 600B MoE model Step 5 Preview and confirmed it can run on 8 H200s.
Weekly AI Papers: AgentBaton Cuts API Costs by 65.9%, Criticisms of MCP Governance Monopoly
A roundup of 10 latest AI/ML papers addressing agent efficiency and MCP governance issues.
Anthropic Donates Agent Skills Standard to AAIF
Anthropic donates the Agent Skills standard to the AAIF, initiating neutral governance.
OpenSandbox Joins AAIF Project Under Linux Foundation
OpenSandbox has been incorporated into the AAIF project under the Linux Foundation to co-manage agent sandbox infrastructure standards.
NVIDIA Nemotron 3.5 Bolt E4B Released for Local Agents with 4GB VRAM
NVIDIA has released Nemotron 3.5 Bolt E4B, a small model for interactive agents that operates on 4GB VRAM.
Anthropic Scion: Checkout Speed 100x Faster Than Git
Anthropic processed 30,000 merges per day and reduced CI CPU usage by 25% by adopting Scion.
OpenEnv Joins AAIF and Releases v0.10.0
OpenEnv joined the AAIF project and released v0.10.0, enabling agent learning environment development to proceed under a neutral foundation.
OpenAI Donates Open Responses Specification to AAIF
OpenAI donated the Open Responses specification to AAIF to promote the standardization of LLM calls.
Modular Open-Sources Mojo 1.0 and MAX Source, Completes Acquisition by Qualcomm
Modular has open-sourced Mojo 1.0 and released the source code for the MAX framework, while completing its acquisition by Qualcomm.
Hugging Face Provides Unified API for Local and Cloud LLMs with AnyLanguageModel 1.0
You can control local and cloud LLMs identically by changing only the import statement.
LiteLLM v1.107.0 Supply Chain Attack, Estimated ~49,000 Downloads
Malicious LiteLLM v1.107.0 was distributed through a manipulated official pipeline, posing a risk of credential theft.
whatsapp-mcp: WhatsApp conversation search MCP server released without sending functionality
Enables read-only search of WhatsApp conversations through self-hosting on Cloudflare Workers.