DeepSeek's new 304B parameter model targets enhanced agentic capabilities and is available for local deployment.
Google's new vision-language-action model adds a high-level reasoning layer to control both tabletop arms and humanoids.
Anthropic officially outlines its policy and safety-focused stance on the release and regulation of open-weights models.
The MCP 2.0 specification introduces stateless operations, simplifying client-server architectures for LLM tool integration.
The update introduces a mechanism allowing agent tools to execute code directly within the user's browser.
A lightweight, open-source evaluation suite designed for rapid testing of models, prompts, and agent harnesses.
Nvidia details attention mechanism co-design strategies to mitigate the high computational cost of long-context agentic inference.
Analysis of how LLMs learn to modulate their reasoning steps for low-, medium-, and high-effort tasks.
An open-weight alternative framework claiming high-quality narrative generation at a fraction of proprietary API costs.
Google adds Gemini 3.6 Flash support and event-driven hooks to its managed agent infrastructure.
A field report detailing how researchers leverage AI coding agents to refactor and modernize legacy scientific computing codebases.
Google retracted its Earth AI generation tool immediately after launch due to severe misinformation and deepfake concerns.