Meituan open-weights LongCat-2.5: 1.6T MoE, vision, MIT-licensed

Weights are live on Hugging Face, hosted API at $0.30/$1.20 per Mtok — while a Fireworks breach and new Claude Code leak figures leave builders with cleanup.

Nowline NOV 10 2:00 PM banner

Top AI stories from the last hour

Top AI stories from the last hour

Copy markdown

  • A 1.6T-param MoE, MIT-licensed, with vision

    Meituan published LongCat-2.5's weights — ~1.6T total, ~48B active per token — under a commercial-friendly MIT license, with image and video input and strong agentic-coding scores (Terminal-Bench v2.1 88.4, DeepSWE 64.6). Pull them from meituan-longcat/LongCat-2.5 on Hugging Face, or skip the GPUs and hit the hosted API at $0.30/$1.20 per Mtok (cached input $0.006).

  • Fireworks AI got breached — rotate your keys

    Fireworks notified customers on Nov 9 that an intruder reached its dev machines on Oct 24 — reportedly its second breach in weeks. Billing data was exposed for everyone, and some accounts also had secrets, datasets, and incomplete model weights taken; if you serve models on Fireworks, rotate API keys and review account activity now.

  • Update: Claude Code's secret leak, quantified

    Security firm Apiiro reports it found 2,539 distinct secrets exposed across public GitHub repos via Claude Code's "Don't Ask Again" file, which stores approved commands — credentials included — in plaintext. If you've ever clicked it, audit your .claude.json and any committed config, and rotate anything that leaked.