Liquid AI releases LFM3: 6 open-weight 2B-class multimodal models

Limits apply to ChatGPT Go, Plus, Pro 100, Pro 200 and Pro 500; free and Business, Enterprise and Edu plans are exempt. Users hitting the cap see "Device limit reached," and logging out does not free a slot.

DeepSeek published an open-source prototype of a more autonomous coding agent on GitHub, aimed at Anthropic's Claude Code. The release comes from DeepSeek's Harness team, which posted the underlying architecture for an always-on "super-agent."

LongCat-2.5 has 1.6 trillion total parameters with roughly 48 billion activated per token, trained on long task sequences and large-scale RL for complex codebases. It handles visual and text inputs in a unified token framework for visual coding, design, and 3D tasks.

The robot debuted at the opening of Guangdong's 2,000th 7-Eleven location on October 29, inside Tencent's Shenzhen headquarters campus. It greets customers, recommends products and bundles, and takes orders by conversation; shoppers scan a code to pay themselves.

The ~9B-parameter multimodal model runs a shared block of 24 decoder layers twice, deepening computation without duplicating weights. Training supports sequences up to 262,144 tokens across text, images, code, and tool-use; weights ship under Apache 2.0.

Beijing-based GenRobot AI closed a $100 million Series A+ in Q3, disclosed November 2, with SAIC Motor, Baidu, Eastworlds, Shoucheng Holdings and Galaxy General Robotics among investors. Its Matrix platform syncs, cleans and labels human demonstration recordings for robot training and evaluation.