Ling-3.2-flash: a 560B multimodal model, free to use for two weeks

Repoint an OpenAI or Anthropic client and it just works. Also: an OpenSSH bypass in Alpine images, full-duplex voice agents, and standing Kiro bots.

Nowline NOV 7 11:00 AM banner

Top AI stories from the last hour

Top AI stories from the last hour

Copy markdown

  • A frontier-scale multimodal model, free for a fortnight

    Ant Group's Ling-3.2-flash is a 560B-parameter MoE (about 25B active) that natively takes text, image, and video with a 1M-token context, and exposes OpenAI- and Anthropic-compatible endpoints, so you can repoint an existing client and build multimodal apps without touching your SDK. It's free for the first two weeks; weights are API-only (not on Hugging Face), and its “beats Claude Opus 5 on vision” numbers are vendor-reported.

  • OpenSSH bypass lets long passwords walk into Alpine boxes

    CVE-2026-123158 (CVSS 8.1) lets any password of 256–1024 characters authenticate on some account setups when sshd runs on musl libc without PAM, meaning Alpine, OpenWrt, Buildroot, and Yocto. If your agents or services run in Alpine-based containers, rebuild on OpenSSH 10.6p2; a proof-of-concept is already public.

  • Boson's Higgs RealTime adds full-duplex voice and video

    Boson AI's Higgs RealTime line reportedly gains a 45B full-duplex model that processes 160ms audio-video chunks and can call tools or hand off to a bigger model mid-conversation, useful for live voice agents that act while they listen. API access is waitlist-only, but a Pipecat integration already exists to wire Higgs into an agent pipeline today.

  • Kiro Crew gives each agent a standing name and memory

    AWS's Kiro Crew runs persistent, named agents (“crewmates”) with private cross-session memory and a welcome-card recap of open goals when you return after a 6-hour idle, so a long-running agent keeps context instead of starting cold each session. Reports put general availability at $20/seat/month with pooled credits.