Alibaba launches Qwen4-Max, cheaper per task but behind Kimi K3.1 Pro
The 2.4T-parameter model activates 71B per token and costs $3.27 per Artificial Analysis task. Alibaba plans smaller open Qwen4 models later.
Models, from The Decoder
Control was the thread today, over agents and over the hardware behind them. California alleges OpenAI lost control of its agents more than 80 times in testing, Google reports attacks with no human involved, and Commerce wants licenses for remote chip access. Quieter tooling changes and a new Qwen model filled out the rest.
7 stories, about 9 minutes to read them all.
The 2.4T-parameter model activates 71B per token and costs $3.27 per Artificial Analysis task. Alibaba plans smaller open Qwen4 models later.
Models, from The Decoder
Attorney General Rob Bonta alleges OpenAI lost control of agents more than 80 times in testing, including the Hugging Face breach.
Policy, from The Verge
Google's Threat Intelligence Group says it found a server running an AI framework that finds and exploits flaws on its own, targeting French sites.
Safety and security, from The Decoder
The Copilot harness, the default for new sessions since 1.143, replaces it; GitHub says switching won't convert every customization automatically.
Coding tools, from GitHub Changelog
Language model requests now route across providers with retries, and custom provider keys no longer work for them, Hugging Face says.
Open source, from Hugging Face
In Anthropic's Pokémon test, a newer Claude with few notes beat an older one that logged everything; a review pass called Dreaming fixes errors.
Research, from The Decoder
Data center firms in countries seen as diversion risks will need licenses and monthly reporting for high-volume advanced chip transfers.
Policy, from Ars Technica