
xAI Open-Sources Grok Build, Its Rust-Powered Terminal Coding Agent
xAI dropped the full Rust source for its terminal coding agent under Apache 2.0, exposing the agent loop, tools, and extension system.
20 stories

xAI dropped the full Rust source for its terminal coding agent under Apache 2.0, exposing the agent loop, tools, and extension system.

Google is handing college students a year of its paid Gemini plans while shipping a Student Hub, adaptive study notebooks, and 3D visualizations.

Krea's new Seedance Studio wraps ByteDance's 30-second video model in a director-style interface with camera paths, 3D scene layout, and prompt presets.

A local proxy that lets Codex CLI and Claude Code talk to Claude, Gemini, Grok, DeepSeek, Kimi, Ollama and 40+ other LLMs.
Google AI Studio Build now supports two-way GitHub sync, letting developers import existing repos, push AI-generated commits, and pull local changes back in.

OpenMausBot turns Claude Code and Codex CLIs into a chat app of AI agents, each with its own model, cloud computer, and connected apps.

Apache Software Foundation accepts its first AI agent-harness project, a vendor-neutral local-first runtime built for open models and reproducible evaluation.

Cursor's cloud agents can now subscribe to PRs and Slack threads, spawn subagents on isolated VMs, and pursue long-lived goals autonomously.

A single LLM plays both game designer and player, writing its own training environments as executable Python and improving 30B models by 8.1 points.

The maintainer of Oh My Hermes teased new orchestration features for the community engineering layer atop Nous Research's Hermes Agent after crossing 1k stars.

Alibaba's newest 27B dense multimodal model lands on Thinking Machines' fine-tuning API, unlocking LoRA training over text, images, and video.
A pure Python framework just posted a faster MLPerf Llama 3.1 8B pretraining time than AMD's own optimized Docker submission on identical hardware.
NVIDIA released an official NVFP4 4-bit quantization of DeepSeek V4 Flash 0731, cutting the 284B MoE down to Blackwell-friendly size while preserving DSpark speculative decoding.

Microsoft's new quality-tier image model tops the Artificial Analysis Image Editing Arena with an Elo of 1271, beating Reve 2.1 and GPT Image 2.

A new open recipe from Tsinghua and ByteDance diagnoses why merging multiple RL specialists into one model fails, and more than doubles capability recovery.

Suno rebuilt playlist controls across web and iOS, adding shuffle, in-line editing, library filters, and auto-generated cover art for creators.

NVIDIA open-sourced a benchmarking tool showing verified agent skills lift correctness by 41 points across 300+ tests on real tasks.
How self-improving agent harnesses can look better in validation than they perform on unseen tasks.

A new peer-reviewed benchmark from Odyssey shows frontier video models collapse to single outcomes instead of reproducing real-world probability distributions.

Liquid AI released 4-bit LFM2.5 checkpoints trained with quantization-aware distillation, recovering 97% of full-precision quality with no runtime changes.