NVIDIA Expands Cosmos for World Models
NVIDIA expanded the Cosmos platform with new training tools, simulation checkpoints, and data factory blueprints to help developers build world models for physical AI systems. The platform ranks in the top five on six open model leaderboards, strengthening its position in the physical AI sector.
Summarized by AI from one outlet
Source
- NVIDIA BlogRead original ↗
Latest
- Open source
Tsinghua Open-Sources AKV LLM Cache Engine
Tsinghua University researchers released AKV, an open-source cache engineering serving system based on SGLang. By dropping and repositioning KV cache instead of summarizing history, it cuts the average cost by 27.7% across four agent benchmarks. This helps optimize execution efficiency and cost for large language models in agent applications.
- Open source
Tencent Releases Open-Source World Model Hy-Nomas
Tencent released the open-source world model Hy-Nomas, which generates game-like interactive virtual worlds and supports local execution on RTX 4090.
- Open source
HKUST and Yanshan Tech open-source new KV cache optimization model
HKUST and Yanshan Tech open-sourced an AI model that compresses KV Cache by 99%. This technology boosts single-card throughput by six times, supporting 512-way parallel long inference.
- Open source
LiteLLM Reveals Supply Chain Attack: Unrevoked Key Compromised Releases
Open-source AI tool LiteLLM released an investigation showing attackers exploited a May-omitted and unrevoked key to gain GitHub admin access and steal maintainer credentials. Hackers then hijacked the release process, pushing malicious versions to PyPI and container repositories, affecting related developers.
- Open source
Nanbeige Releases Open-Source Agentic Coding Model 4.3-33B
Nanbeige LLM Lab released Nanbeige4.3-33B, a 33B parameter model licensed under Apache 2.0 with 512K context support. The model scored 80.9 on Terminal-Bench 2.1 and will be used to teach the upcoming smaller 3B variant.
All news → · All rights belong to the original creators