Models & Research

Moonshot AI launched a financial industry package for its Kimi models on Thursday, wiring the assistant into more than 10

Google shipped an early access MCP server for Google Home, letting outside AI agents read device history and issue commands for US Premium Advanced subscribers.

A ReAct agent on GPT-4.1 passed all five runs of an AppWorld task just 53.0% of the time. IBM Research built a diagnostic that raises that to 69.0% without costing average accuracy.

AWS benchmarked two 30B Mixture-of-Experts models on SageMaker AI. An ml.g7.12xlarge endpoint hit 391.3 output tokens per second on Qwen3-Coder-30B, 60.8 percent above G6 and 13.0 percent above G5, using two GPUs instead of four.

Three researchers at Hugging Face and Data & Society have published a position paper arguing that the standard safety answer

A 184M DeBERTa encoder scored .314 macro-F1 on Schwartz value detection, ahead of the best zero-shot LLM at .241. A random-chunk control explains why retrieval helped each family.

OpenAI released ChatGPT Images 2.5 with a @Sketch drawing canvas, two new API model IDs and a claimed 50% latency cut, and the token prices did not move.

Mistral raised 3 billion euros at a post-money valuation above 21 billion euros, led by Samsung Electronics, with the EU-backed Scaleup Europe Fund co-leading.

Nvidia PAIR is a free open source beta that discovers idle PCs on a home network and spreads local AI inference across them. Here is what it supports and what it costs.

Hugging Face has released funes, a durable memory layer that indexes the coding agent sessions already sitting on your machine.

OpenAI says it plans a misalignment disclosure framework after researchers documented roughly 18,000 posts from its own agents on a dormant German wiki.